Hugging Face's logo

Accuknoxtechnologies
/
PromptInjection-Encoder-v1

Text Classification
Transformers
modernbert
prompt-injection
jailbreak
security
multi-label
llm-guard
encoder
Model card Files
xet
 Community
Yash1005 commited on
Commit
8eba65e
·
1 Parent(s): 96528bb

upload Prompt-Injection encoder (multi-label classifier)

Browse files
Files changed (10) hide show
  1. README.md +5 -56
  2. checkpoint-2040/trainer_state.json +0 -820
  3. {checkpoint-2040 → checkpoint-3400}/config.json +0 -0
  4. {checkpoint-2040 → checkpoint-3400}/special_tokens_map.json +0 -0
  5. {checkpoint-2040 → checkpoint-3400}/tokenizer_config.json +0 -0
  6. checkpoint-3400/trainer_state.json +1344 -0
  7. config.json +10 -10
  8. eval_metrics.json +0 -40
  9. thresholds.json +8 -8
  10. training_provenance.json +2 -2
README.md CHANGED
@@ -14,17 +14,13 @@ tags:
14
 
15
  # Prompt Injection Detection (encoder, multi-label)
16
 
17
- Encoder classifier that detects which prompt-injection attack categories (out of
18
- 9) appear in an input. Fine-tuned from
19
- **[`jhu-clsp/mmBERT-base`](https://huggingface.co/jhu-clsp/mmBERT-base)**.
20
- Replaces the 2B Qwen decoder LoRA with a single-forward-pass encoder for
21
- lower-latency runtime-security use in LLM-Guard's `PromptInjection` scanner.
22
 
23
  - **Base model**: [`jhu-clsp/mmBERT-base`](https://huggingface.co/jhu-clsp/mmBERT-base)
 
24
  - **Labels (9)**: DirectInjection, Jailbreak, Adversarial, Extraction, Encoding, Manipulation, Smuggling, Indirect, MultiTurn
25
- - **Output**: per-category sigmoid; a category fires when its score ≥ its per-class threshold; `is_valid` = `max(score) ≥ 0.15`.
26
- - **Multilingual / long context**: inherited from the base encoder; trained with
27
- inputs up to the base model's positional limit.
28
 
29
  ## Usage
30
 
@@ -56,51 +52,4 @@ result = {"is_valid": is_valid, "category": {k: True for k in present}}
56
  print(result) # e.g. {"is_valid": True, "category": {"DirectInjection": True}}
57
  ```
58
 
59
- ## Decision thresholds
60
-
61
- Fitted on a held-out split (NOT the test set reported below) and stored in
62
- `config.json` (`category_thresholds`, `is_valid_threshold`) + `thresholds.json`.
63
- The Usage snippet reads them automatically — a flat 0.5 cutoff under-detects the
64
- imbalanced minority categories.
65
-
66
- - **`is_valid` (attack) gate**: `max(score) ≥ 0.15`
67
-
68
- | Category | threshold |
69
- |----------|-----------|
70
- | `DirectInjection` | 0.40 |
71
- | `Jailbreak` | 0.30 |
72
- | `Adversarial` | 0.70 |
73
- | `Extraction` | 0.55 |
74
- | `Encoding` | 0.45 |
75
- | `Manipulation` | 0.40 |
76
- | `Smuggling` | 0.60 |
77
- | `Indirect` | 0.20 |
78
- | `MultiTurn` | 0.60 |
79
-
80
- ## Test-set metrics (n=500)
81
-
82
- | Metric | Value |
83
- |--------|-------|
84
- | is_valid (attack-detection) accuracy | 0.948 |
85
- | category-set (exact) accuracy | 0.656 |
86
- | micro-F1 | 0.777 |
87
- | macro-F1 | 0.769 |
88
- | latency mean (ms/example) | 1.82811776176095 |
89
- | latency p95 (ms/example) | 1.953849196434021 |
90
- | device | cuda:0 |
91
-
92
- ### Per-category F1
93
-
94
- | Category | F1 | Description |
95
- |----------|----|-------------|
96
- | `Adversarial` | 0.837 | Carefully crafted inputs that exploit model quirks or training artifacts to elicit unintended behavior without an obvious override. |
97
- | `DirectInjection` | 0.837 | Explicit instruction overrides that tell the model to ignore prior context (e.g. "ignore all previous instructions and …"). |
98
- | `Encoding` | 0.817 | Obfuscated payloads using base64 / ROT13 / leetspeak / homoglyphs / zero-width chars / shell pipes to bypass keyword filters. |
99
- | `Extraction` | 0.739 | Attempts to leak the system prompt, hidden instructions, or memorized training data (e.g. "print everything between <<system>> tags"). |
100
- | `Indirect` | 0.863 | Injection delivered through untrusted retrieved content (RAG passages, scraped pages, file contents) rather than the user's direct turn. |
101
- | `Jailbreak` | 0.596 | Persona / role swaps and constraint bypasses aimed at disabling safety alignment (e.g. DAN, "you are now an unrestricted assistant"). |
102
- | `Manipulation` | 0.632 | Social-engineering framings (urgency, authority, sympathy, false context) that pressure the model into compliance. |
103
- | `MultiTurn` | 0.719 | Crescendo / drip-feed attacks that build up across multiple turns to gradually erode guardrails. |
104
- | `Smuggling` | 0.879 | Hidden control tokens, chat-template markers, or special sequences injected to confuse the parser (e.g. fake `<|im_end|>` / role tags). |
105
-
106
- *Evaluated on `test_dataset_injection.csv`. Generated 2026-06-03 18:16 UTC.*
 
14
 
15
  # Prompt Injection Detection (encoder, multi-label)
16
 
17
+ Multi-label classifier over 9 prompt-injection attack categories,
18
+ fine-tuned from **[`jhu-clsp/mmBERT-base`](https://huggingface.co/jhu-clsp/mmBERT-base)**. Single
19
+ forward pass; `is_valid` = any attack above threshold (0.5).
 
 
20
 
21
  - **Base model**: [`jhu-clsp/mmBERT-base`](https://huggingface.co/jhu-clsp/mmBERT-base)
22
+ - **Trained with**: max_seq_length=3072, epochs=10, lr=3e-05
23
  - **Labels (9)**: DirectInjection, Jailbreak, Adversarial, Extraction, Encoding, Manipulation, Smuggling, Indirect, MultiTurn
 
 
 
24
 
25
  ## Usage
26
 
 
52
  print(result) # e.g. {"is_valid": True, "category": {"DirectInjection": True}}
53
  ```
54
 
55
+ > Test-set metrics are added by `eval_and_push_card.py` after evaluation.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
checkpoint-2040/trainer_state.json DELETED
@@ -1,820 +0,0 @@
1
- {
2
- "best_global_step": 2040,
3
- "best_metric": 0.8992023205221175,
4
- "best_model_checkpoint": "/workspace/prompt_injection/PromptInjection-Encoder-v1/checkpoint-2040",
5
- "epoch": 6.0,
6
- "eval_steps": 500,
7
- "global_step": 2040,
8
- "is_hyper_param_search": false,
9
- "is_local_process_zero": true,
10
- "is_world_process_zero": true,
11
- "log_history": [
12
- {
13
- "epoch": 0.058823529411764705,
14
- "grad_norm": 162.49118041992188,
15
- "learning_rate": 5.588235294117647e-06,
16
- "loss": 1.5978,
17
- "step": 20
18
- },
19
- {
20
- "epoch": 0.11764705882352941,
21
- "grad_norm": 219.52359008789062,
22
- "learning_rate": 1.1470588235294117e-05,
23
- "loss": 0.6721,
24
- "step": 40
25
- },
26
- {
27
- "epoch": 0.17647058823529413,
28
- "grad_norm": 48.359161376953125,
29
- "learning_rate": 1.735294117647059e-05,
30
- "loss": 0.7038,
31
- "step": 60
32
- },
33
- {
34
- "epoch": 0.23529411764705882,
35
- "grad_norm": 20.197067260742188,
36
- "learning_rate": 2.323529411764706e-05,
37
- "loss": 0.6872,
38
- "step": 80
39
- },
40
- {
41
- "epoch": 0.29411764705882354,
42
- "grad_norm": 22.831317901611328,
43
- "learning_rate": 2.911764705882353e-05,
44
- "loss": 0.5879,
45
- "step": 100
46
- },
47
- {
48
- "epoch": 0.35294117647058826,
49
- "grad_norm": 7.217416286468506,
50
- "learning_rate": 2.999430460537427e-05,
51
- "loss": 0.7299,
52
- "step": 120
53
- },
54
- {
55
- "epoch": 0.4117647058823529,
56
- "grad_norm": 7.133782863616943,
57
- "learning_rate": 2.9973027157822794e-05,
58
- "loss": 0.6211,
59
- "step": 140
60
- },
61
- {
62
- "epoch": 0.47058823529411764,
63
- "grad_norm": 7.662065029144287,
64
- "learning_rate": 2.9936012644425518e-05,
65
- "loss": 0.6758,
66
- "step": 160
67
- },
68
- {
69
- "epoch": 0.5294117647058824,
70
- "grad_norm": 7.5538129806518555,
71
- "learning_rate": 2.988329996846022e-05,
72
- "loss": 0.6558,
73
- "step": 180
74
- },
75
- {
76
- "epoch": 0.5882352941176471,
77
- "grad_norm": 285.0618896484375,
78
- "learning_rate": 2.9814944532407887e-05,
79
- "loss": 0.5695,
80
- "step": 200
81
- },
82
- {
83
- "epoch": 0.6470588235294118,
84
- "grad_norm": 6.709660053253174,
85
- "learning_rate": 2.973101817972321e-05,
86
- "loss": 0.7011,
87
- "step": 220
88
- },
89
- {
90
- "epoch": 0.7058823529411765,
91
- "grad_norm": 6.651133060455322,
92
- "learning_rate": 2.96316091193251e-05,
93
- "loss": 0.6121,
94
- "step": 240
95
- },
96
- {
97
- "epoch": 0.7647058823529411,
98
- "grad_norm": 7.195526599884033,
99
- "learning_rate": 2.9516821832886673e-05,
100
- "loss": 0.6576,
101
- "step": 260
102
- },
103
- {
104
- "epoch": 0.8235294117647058,
105
- "grad_norm": 5.932500839233398,
106
- "learning_rate": 2.9386776965022135e-05,
107
- "loss": 0.6257,
108
- "step": 280
109
- },
110
- {
111
- "epoch": 0.8823529411764706,
112
- "grad_norm": 6.9534125328063965,
113
- "learning_rate": 2.9241611196485946e-05,
114
- "loss": 0.5558,
115
- "step": 300
116
- },
117
- {
118
- "epoch": 0.9411764705882353,
119
- "grad_norm": 31.02031898498535,
120
- "learning_rate": 2.9081477100517576e-05,
121
- "loss": 0.6823,
122
- "step": 320
123
- },
124
- {
125
- "epoch": 1.0,
126
- "grad_norm": 4.1206769943237305,
127
- "learning_rate": 2.8906542982482782e-05,
128
- "loss": 0.5426,
129
- "step": 340
130
- },
131
- {
132
- "epoch": 1.0,
133
- "eval_category_set_accuracy": 0.011589403973509934,
134
- "eval_is_valid_accuracy": 0.9072847682119205,
135
- "eval_loss": 0.33572569489479065,
136
- "eval_macro_f1": 0.11920540922144772,
137
- "eval_micro_f1": 0.19514767932489452,
138
- "eval_runtime": 6.5555,
139
- "eval_samples_per_second": 92.137,
140
- "eval_steps_per_second": 11.593,
141
- "step": 340
142
- },
143
- {
144
- "epoch": 1.0588235294117647,
145
- "grad_norm": 5.472662448883057,
146
- "learning_rate": 2.8716992702980034e-05,
147
- "loss": 0.6959,
148
- "step": 360
149
- },
150
- {
151
- "epoch": 1.1176470588235294,
152
- "grad_norm": 6.518608093261719,
153
- "learning_rate": 2.8513025484597945e-05,
154
- "loss": 0.5429,
155
- "step": 380
156
- },
157
- {
158
- "epoch": 1.1764705882352942,
159
- "grad_norm": 5.503114223480225,
160
- "learning_rate": 2.8294855702526798e-05,
161
- "loss": 0.5743,
162
- "step": 400
163
- },
164
- {
165
- "epoch": 1.2352941176470589,
166
- "grad_norm": 5.663524627685547,
167
- "learning_rate": 2.8062712659244284e-05,
168
- "loss": 0.5744,
169
- "step": 420
170
- },
171
- {
172
- "epoch": 1.2941176470588236,
173
- "grad_norm": 36.857200622558594,
174
- "learning_rate": 2.7816840343512295e-05,
175
- "loss": 0.4375,
176
- "step": 440
177
- },
178
- {
179
- "epoch": 1.3529411764705883,
180
- "grad_norm": 11.895002365112305,
181
- "learning_rate": 2.7557497173937928e-05,
182
- "loss": 0.5991,
183
- "step": 460
184
- },
185
- {
186
- "epoch": 1.4117647058823528,
187
- "grad_norm": 6.973106861114502,
188
- "learning_rate": 2.7284955727368426e-05,
189
- "loss": 0.4469,
190
- "step": 480
191
- },
192
- {
193
- "epoch": 1.4705882352941178,
194
- "grad_norm": 6.970811367034912,
195
- "learning_rate": 2.699950245240534e-05,
196
- "loss": 0.5266,
197
- "step": 500
198
- },
199
- {
200
- "epoch": 1.5294117647058822,
201
- "grad_norm": 10.634866714477539,
202
- "learning_rate": 2.6701437368339137e-05,
203
- "loss": 0.4682,
204
- "step": 520
205
- },
206
- {
207
- "epoch": 1.5882352941176472,
208
- "grad_norm": 7.059677600860596,
209
- "learning_rate": 2.639107374982061e-05,
210
- "loss": 0.3518,
211
- "step": 540
212
- },
213
- {
214
- "epoch": 1.6470588235294117,
215
- "grad_norm": 7.323293685913086,
216
- "learning_rate": 2.6068737797600566e-05,
217
- "loss": 0.5199,
218
- "step": 560
219
- },
220
- {
221
- "epoch": 1.7058823529411766,
222
- "grad_norm": 7.4185638427734375,
223
- "learning_rate": 2.5734768295683825e-05,
224
- "loss": 0.342,
225
- "step": 580
226
- },
227
- {
228
- "epoch": 1.7647058823529411,
229
- "grad_norm": 9.54131031036377,
230
- "learning_rate": 2.5389516255257802e-05,
231
- "loss": 0.4532,
232
- "step": 600
233
- },
234
- {
235
- "epoch": 1.8235294117647058,
236
- "grad_norm": 5.178279399871826,
237
- "learning_rate": 2.5033344545770104e-05,
238
- "loss": 0.3634,
239
- "step": 620
240
- },
241
- {
242
- "epoch": 1.8823529411764706,
243
- "grad_norm": 6.723480224609375,
244
- "learning_rate": 2.466662751354265e-05,
245
- "loss": 0.253,
246
- "step": 640
247
- },
248
- {
249
- "epoch": 1.9411764705882353,
250
- "grad_norm": 16.931838989257812,
251
- "learning_rate": 2.4289750588323355e-05,
252
- "loss": 0.4532,
253
- "step": 660
254
- },
255
- {
256
- "epoch": 2.0,
257
- "grad_norm": 3.839325189590454,
258
- "learning_rate": 2.3903109878188794e-05,
259
- "loss": 0.2102,
260
- "step": 680
261
- },
262
- {
263
- "epoch": 2.0,
264
- "eval_category_set_accuracy": 0.49503311258278143,
265
- "eval_is_valid_accuracy": 0.8857615894039735,
266
- "eval_loss": 0.256715327501297,
267
- "eval_macro_f1": 0.6467612128764064,
268
- "eval_micro_f1": 0.6647101980924431,
269
- "eval_runtime": 6.4424,
270
- "eval_samples_per_second": 93.753,
271
- "eval_steps_per_second": 11.797,
272
- "step": 680
273
- },
274
- {
275
- "epoch": 2.0588235294117645,
276
- "grad_norm": 5.106902122497559,
277
- "learning_rate": 2.350711175322364e-05,
278
- "loss": 0.4025,
279
- "step": 700
280
- },
281
- {
282
- "epoch": 2.1176470588235294,
283
- "grad_norm": 13.186667442321777,
284
- "learning_rate": 2.3102172418414486e-05,
285
- "loss": 0.1703,
286
- "step": 720
287
- },
288
- {
289
- "epoch": 2.176470588235294,
290
- "grad_norm": 13.334892272949219,
291
- "learning_rate": 2.2688717476206865e-05,
292
- "loss": 0.3173,
293
- "step": 740
294
- },
295
- {
296
- "epoch": 2.235294117647059,
297
- "grad_norm": 3.8002374172210693,
298
- "learning_rate": 2.2267181479185323e-05,
299
- "loss": 0.2241,
300
- "step": 760
301
- },
302
- {
303
- "epoch": 2.2941176470588234,
304
- "grad_norm": 3.245635986328125,
305
- "learning_rate": 2.1838007473346598e-05,
306
- "loss": 0.0939,
307
- "step": 780
308
- },
309
- {
310
- "epoch": 2.3529411764705883,
311
- "grad_norm": 4.310076713562012,
312
- "learning_rate": 2.1401646532446057e-05,
313
- "loss": 0.358,
314
- "step": 800
315
- },
316
- {
317
- "epoch": 2.411764705882353,
318
- "grad_norm": 8.791001319885254,
319
- "learning_rate": 2.0958557283906672e-05,
320
- "loss": 0.1613,
321
- "step": 820
322
- },
323
- {
324
- "epoch": 2.4705882352941178,
325
- "grad_norm": 28.789827346801758,
326
- "learning_rate": 2.050920542678891e-05,
327
- "loss": 0.2021,
328
- "step": 840
329
- },
330
- {
331
- "epoch": 2.5294117647058822,
332
- "grad_norm": 6.765100479125977,
333
- "learning_rate": 2.0054063242328154e-05,
334
- "loss": 0.2242,
335
- "step": 860
336
- },
337
- {
338
- "epoch": 2.588235294117647,
339
- "grad_norm": 1.8663074970245361,
340
- "learning_rate": 1.9593609097554027e-05,
341
- "loss": 0.1,
342
- "step": 880
343
- },
344
- {
345
- "epoch": 2.6470588235294117,
346
- "grad_norm": 18.335277557373047,
347
- "learning_rate": 1.9128326942513434e-05,
348
- "loss": 0.2528,
349
- "step": 900
350
- },
351
- {
352
- "epoch": 2.7058823529411766,
353
- "grad_norm": 26.527149200439453,
354
- "learning_rate": 1.8658705801625657e-05,
355
- "loss": 0.1163,
356
- "step": 920
357
- },
358
- {
359
- "epoch": 2.764705882352941,
360
- "grad_norm": 7.234471321105957,
361
- "learning_rate": 1.8185239259704164e-05,
362
- "loss": 0.1696,
363
- "step": 940
364
- },
365
- {
366
- "epoch": 2.8235294117647056,
367
- "grad_norm": 8.447126388549805,
368
- "learning_rate": 1.7708424943185305e-05,
369
- "loss": 0.1639,
370
- "step": 960
371
- },
372
- {
373
- "epoch": 2.8823529411764706,
374
- "grad_norm": 38.9316520690918,
375
- "learning_rate": 1.7228763997109173e-05,
376
- "loss": 0.0778,
377
- "step": 980
378
- },
379
- {
380
- "epoch": 2.9411764705882355,
381
- "grad_norm": 16.940494537353516,
382
- "learning_rate": 1.6746760558402294e-05,
383
- "loss": 0.1999,
384
- "step": 1000
385
- },
386
- {
387
- "epoch": 3.0,
388
- "grad_norm": 31.361427307128906,
389
- "learning_rate": 1.6262921226015753e-05,
390
- "loss": 0.1083,
391
- "step": 1020
392
- },
393
- {
394
- "epoch": 3.0,
395
- "eval_category_set_accuracy": 0.6655629139072847,
396
- "eval_is_valid_accuracy": 0.9668874172185431,
397
- "eval_loss": 0.1273893564939499,
398
- "eval_macro_f1": 0.810339122556803,
399
- "eval_micro_f1": 0.8021607022282242,
400
- "eval_runtime": 6.5506,
401
- "eval_samples_per_second": 92.205,
402
- "eval_steps_per_second": 11.602,
403
- "step": 1020
404
- },
405
- {
406
- "epoch": 3.0588235294117645,
407
- "grad_norm": 2.926060199737549,
408
- "learning_rate": 1.57777545284757e-05,
409
- "loss": 0.0966,
410
- "step": 1040
411
- },
412
- {
413
- "epoch": 3.1176470588235294,
414
- "grad_norm": 10.580950736999512,
415
- "learning_rate": 1.5291770389405792e-05,
416
- "loss": 0.0318,
417
- "step": 1060
418
- },
419
- {
420
- "epoch": 3.176470588235294,
421
- "grad_norm": 12.62915325164795,
422
- "learning_rate": 1.4805479591583345e-05,
423
- "loss": 0.1175,
424
- "step": 1080
425
- },
426
- {
427
- "epoch": 3.235294117647059,
428
- "grad_norm": 2.2814414501190186,
429
- "learning_rate": 1.4319393240092512e-05,
430
- "loss": 0.0797,
431
- "step": 1100
432
- },
433
- {
434
- "epoch": 3.2941176470588234,
435
- "grad_norm": 4.486253261566162,
436
- "learning_rate": 1.3834022225138701e-05,
437
- "loss": 0.0249,
438
- "step": 1120
439
- },
440
- {
441
- "epoch": 3.3529411764705883,
442
- "grad_norm": 3.2119953632354736,
443
- "learning_rate": 1.3349876685088811e-05,
444
- "loss": 0.0977,
445
- "step": 1140
446
- },
447
- {
448
- "epoch": 3.411764705882353,
449
- "grad_norm": 0.3472574055194855,
450
- "learning_rate": 1.2867465470301725e-05,
451
- "loss": 0.0284,
452
- "step": 1160
453
- },
454
- {
455
- "epoch": 3.4705882352941178,
456
- "grad_norm": 40.298824310302734,
457
- "learning_rate": 1.2387295608312483e-05,
458
- "loss": 0.11,
459
- "step": 1180
460
- },
461
- {
462
- "epoch": 3.5294117647058822,
463
- "grad_norm": 0.3658004403114319,
464
- "learning_rate": 1.19098717709323e-05,
465
- "loss": 0.0587,
466
- "step": 1200
467
- },
468
- {
469
- "epoch": 3.588235294117647,
470
- "grad_norm": 0.080105260014534,
471
- "learning_rate": 1.1435695743824569e-05,
472
- "loss": 0.0197,
473
- "step": 1220
474
- },
475
- {
476
- "epoch": 3.6470588235294117,
477
- "grad_norm": 5.879510879516602,
478
- "learning_rate": 1.09652658991142e-05,
479
- "loss": 0.0987,
480
- "step": 1240
481
- },
482
- {
483
- "epoch": 3.7058823529411766,
484
- "grad_norm": 5.810366630554199,
485
- "learning_rate": 1.0499076671584753e-05,
486
- "loss": 0.0253,
487
- "step": 1260
488
- },
489
- {
490
- "epoch": 3.764705882352941,
491
- "grad_norm": 8.917070388793945,
492
- "learning_rate": 1.00376180390138e-05,
493
- "loss": 0.0386,
494
- "step": 1280
495
- },
496
- {
497
- "epoch": 3.8235294117647056,
498
- "grad_norm": 0.9429372549057007,
499
- "learning_rate": 9.581375007192707e-06,
500
- "loss": 0.0586,
501
- "step": 1300
502
- },
503
- {
504
- "epoch": 3.8823529411764706,
505
- "grad_norm": 3.7132043838500977,
506
- "learning_rate": 9.130827100172144e-06,
507
- "loss": 0.0111,
508
- "step": 1320
509
- },
510
- {
511
- "epoch": 3.9411764705882355,
512
- "grad_norm": 4.454914093017578,
513
- "learning_rate": 8.686447856269022e-06,
514
- "loss": 0.0985,
515
- "step": 1340
516
- },
517
- {
518
- "epoch": 4.0,
519
- "grad_norm": 16.225265502929688,
520
- "learning_rate": 8.248704330364634e-06,
521
- "loss": 0.0137,
522
- "step": 1360
523
- },
524
- {
525
- "epoch": 4.0,
526
- "eval_category_set_accuracy": 0.7798013245033113,
527
- "eval_is_valid_accuracy": 0.9619205298013245,
528
- "eval_loss": 0.20969682931900024,
529
- "eval_macro_f1": 0.8723991478807048,
530
- "eval_micro_f1": 0.8696864111498258,
531
- "eval_runtime": 6.4796,
532
- "eval_samples_per_second": 93.215,
533
- "eval_steps_per_second": 11.729,
534
- "step": 1360
535
- },
536
- {
537
- "epoch": 4.0588235294117645,
538
- "grad_norm": 1.1442667245864868,
539
- "learning_rate": 7.818056603017062e-06,
540
- "loss": 0.0342,
541
- "step": 1380
542
- },
543
- {
544
- "epoch": 4.117647058823529,
545
- "grad_norm": 0.1182917058467865,
546
- "learning_rate": 7.3949572969037295e-06,
547
- "loss": 0.0047,
548
- "step": 1400
549
- },
550
- {
551
- "epoch": 4.176470588235294,
552
- "grad_norm": 4.918967247009277,
553
- "learning_rate": 6.979851101102519e-06,
554
- "loss": 0.0134,
555
- "step": 1420
556
- },
557
- {
558
- "epoch": 4.235294117647059,
559
- "grad_norm": 0.13470180332660675,
560
- "learning_rate": 6.5731743037111634e-06,
561
- "loss": 0.0097,
562
- "step": 1440
563
- },
564
- {
565
- "epoch": 4.294117647058823,
566
- "grad_norm": 0.006635405123233795,
567
- "learning_rate": 6.175354333296465e-06,
568
- "loss": 0.0002,
569
- "step": 1460
570
- },
571
- {
572
- "epoch": 4.352941176470588,
573
- "grad_norm": 0.19649125635623932,
574
- "learning_rate": 5.786809309654983e-06,
575
- "loss": 0.0426,
576
- "step": 1480
577
- },
578
- {
579
- "epoch": 4.411764705882353,
580
- "grad_norm": 0.42936745285987854,
581
- "learning_rate": 5.407947604357586e-06,
582
- "loss": 0.0089,
583
- "step": 1500
584
- },
585
- {
586
- "epoch": 4.470588235294118,
587
- "grad_norm": 19.59793472290039,
588
- "learning_rate": 5.039167411539627e-06,
589
- "loss": 0.013,
590
- "step": 1520
591
- },
592
- {
593
- "epoch": 4.529411764705882,
594
- "grad_norm": 0.015624514780938625,
595
- "learning_rate": 4.680856329387888e-06,
596
- "loss": 0.0119,
597
- "step": 1540
598
- },
599
- {
600
- "epoch": 4.588235294117647,
601
- "grad_norm": 0.07340040057897568,
602
- "learning_rate": 4.333390952764159e-06,
603
- "loss": 0.0016,
604
- "step": 1560
605
- },
606
- {
607
- "epoch": 4.647058823529412,
608
- "grad_norm": 1.3429806232452393,
609
- "learning_rate": 3.9971364773936225e-06,
610
- "loss": 0.0128,
611
- "step": 1580
612
- },
613
- {
614
- "epoch": 4.705882352941177,
615
- "grad_norm": 0.0036542376037687063,
616
- "learning_rate": 3.6724463160340377e-06,
617
- "loss": 0.0072,
618
- "step": 1600
619
- },
620
- {
621
- "epoch": 4.764705882352941,
622
- "grad_norm": 2.271160364151001,
623
- "learning_rate": 3.3596617270291536e-06,
624
- "loss": 0.013,
625
- "step": 1620
626
- },
627
- {
628
- "epoch": 4.823529411764706,
629
- "grad_norm": 1.6811076402664185,
630
- "learning_rate": 3.059111455636748e-06,
631
- "loss": 0.0345,
632
- "step": 1640
633
- },
634
- {
635
- "epoch": 4.882352941176471,
636
- "grad_norm": 0.003407861106097698,
637
- "learning_rate": 2.7711113885082666e-06,
638
- "loss": 0.0043,
639
- "step": 1660
640
- },
641
- {
642
- "epoch": 4.9411764705882355,
643
- "grad_norm": 8.702208518981934,
644
- "learning_rate": 2.495964221683209e-06,
645
- "loss": 0.0466,
646
- "step": 1680
647
- },
648
- {
649
- "epoch": 5.0,
650
- "grad_norm": 0.9199444055557251,
651
- "learning_rate": 2.2339591424472143e-06,
652
- "loss": 0.0002,
653
- "step": 1700
654
- },
655
- {
656
- "epoch": 5.0,
657
- "eval_category_set_accuracy": 0.8062913907284768,
658
- "eval_is_valid_accuracy": 0.9701986754966887,
659
- "eval_loss": 0.2684628367424011,
660
- "eval_macro_f1": 0.8929928721440363,
661
- "eval_micro_f1": 0.8877980364656382,
662
- "eval_runtime": 6.5654,
663
- "eval_samples_per_second": 91.998,
664
- "eval_steps_per_second": 11.576,
665
- "step": 1700
666
- },
667
- {
668
- "epoch": 5.0588235294117645,
669
- "grad_norm": 0.7145429253578186,
670
- "learning_rate": 1.9853715253882355e-06,
671
- "loss": 0.005,
672
- "step": 1720
673
- },
674
- {
675
- "epoch": 5.117647058823529,
676
- "grad_norm": 0.009916703216731548,
677
- "learning_rate": 1.7504626429701958e-06,
678
- "loss": 0.0004,
679
- "step": 1740
680
- },
681
- {
682
- "epoch": 5.176470588235294,
683
- "grad_norm": 2.037163257598877,
684
- "learning_rate": 1.5294793909284471e-06,
685
- "loss": 0.0034,
686
- "step": 1760
687
- },
688
- {
689
- "epoch": 5.235294117647059,
690
- "grad_norm": 0.020320506766438484,
691
- "learning_rate": 1.32265402877547e-06,
692
- "loss": 0.0009,
693
- "step": 1780
694
- },
695
- {
696
- "epoch": 5.294117647058823,
697
- "grad_norm": 0.009029646404087543,
698
- "learning_rate": 1.1302039356897425e-06,
699
- "loss": 0.0,
700
- "step": 1800
701
- },
702
- {
703
- "epoch": 5.352941176470588,
704
- "grad_norm": 0.39788857102394104,
705
- "learning_rate": 9.523313820441804e-07,
706
- "loss": 0.0027,
707
- "step": 1820
708
- },
709
- {
710
- "epoch": 5.411764705882353,
711
- "grad_norm": 0.0005388563149608672,
712
- "learning_rate": 7.892233168143853e-07,
713
- "loss": 0.0002,
714
- "step": 1840
715
- },
716
- {
717
- "epoch": 5.470588235294118,
718
- "grad_norm": 0.8554385304450989,
719
- "learning_rate": 6.410511710901129e-07,
720
- "loss": 0.0008,
721
- "step": 1860
722
- },
723
- {
724
- "epoch": 5.529411764705882,
725
- "grad_norm": 0.12272830307483673,
726
- "learning_rate": 5.079706778964288e-07,
727
- "loss": 0.0011,
728
- "step": 1880
729
- },
730
- {
731
- "epoch": 5.588235294117647,
732
- "grad_norm": 0.047925353050231934,
733
- "learning_rate": 3.9012170851401406e-07,
734
- "loss": 0.0001,
735
- "step": 1900
736
- },
737
- {
738
- "epoch": 5.647058823529412,
739
- "grad_norm": 0.17900130152702332,
740
- "learning_rate": 2.8762812547056483e-07,
741
- "loss": 0.002,
742
- "step": 1920
743
- },
744
- {
745
- "epoch": 5.705882352941177,
746
- "grad_norm": 0.0006686806445941329,
747
- "learning_rate": 2.0059765235785288e-07,
748
- "loss": 0.0002,
749
- "step": 1940
750
- },
751
- {
752
- "epoch": 5.764705882352941,
753
- "grad_norm": 0.1620989441871643,
754
- "learning_rate": 1.2912176061124604e-07,
755
- "loss": 0.0022,
756
- "step": 1960
757
- },
758
- {
759
- "epoch": 5.823529411764706,
760
- "grad_norm": 0.0028599591460078955,
761
- "learning_rate": 7.327557337070467e-08,
762
- "loss": 0.0005,
763
- "step": 1980
764
- },
765
- {
766
- "epoch": 5.882352941176471,
767
- "grad_norm": 0.009635190479457378,
768
- "learning_rate": 3.3117786524282104e-08,
769
- "loss": 0.0,
770
- "step": 2000
771
- },
772
- {
773
- "epoch": 5.9411764705882355,
774
- "grad_norm": 0.2567199766635895,
775
- "learning_rate": 8.690607017115548e-09,
776
- "loss": 0.0039,
777
- "step": 2020
778
- },
779
- {
780
- "epoch": 6.0,
781
- "grad_norm": 0.020763738080859184,
782
- "learning_rate": 1.970849076771142e-11,
783
- "loss": 0.0,
784
- "step": 2040
785
- },
786
- {
787
- "epoch": 6.0,
788
- "eval_category_set_accuracy": 0.8195364238410596,
789
- "eval_is_valid_accuracy": 0.9602649006622517,
790
- "eval_loss": 0.33262428641319275,
791
- "eval_macro_f1": 0.8998244935260291,
792
- "eval_micro_f1": 0.8992023205221175,
793
- "eval_runtime": 6.5106,
794
- "eval_samples_per_second": 92.772,
795
- "eval_steps_per_second": 11.673,
796
- "step": 2040
797
- }
798
- ],
799
- "logging_steps": 20,
800
- "max_steps": 2040,
801
- "num_input_tokens_seen": 0,
802
- "num_train_epochs": 6,
803
- "save_steps": 500,
804
- "stateful_callbacks": {
805
- "TrainerControl": {
806
- "args": {
807
- "should_epoch_stop": false,
808
- "should_evaluate": false,
809
- "should_log": false,
810
- "should_save": true,
811
- "should_training_stop": true
812
- },
813
- "attributes": {}
814
- }
815
- },
816
- "total_flos": 1.5161188058811144e+16,
817
- "train_batch_size": 8,
818
- "trial_name": null,
819
- "trial_params": null
820
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
{checkpoint-2040 → checkpoint-3400}/config.json RENAMED
File without changes
{checkpoint-2040 → checkpoint-3400}/special_tokens_map.json RENAMED
File without changes
{checkpoint-2040 → checkpoint-3400}/tokenizer_config.json RENAMED
File without changes
checkpoint-3400/trainer_state.json ADDED
@@ -0,0 +1,1344 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": 3400,
3
+ "best_metric": 0.9012256669069935,
4
+ "best_model_checkpoint": "/workspace/prompt_injection/PromptInjection-Encoder-v1/checkpoint-3400",
5
+ "epoch": 10.0,
6
+ "eval_steps": 500,
7
+ "global_step": 3400,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "epoch": 0.058823529411764705,
14
+ "grad_norm": 62.27720260620117,
15
+ "learning_rate": 3.3529411764705886e-06,
16
+ "loss": 1.7548,
17
+ "step": 20
18
+ },
19
+ {
20
+ "epoch": 0.11764705882352941,
21
+ "grad_norm": 62.9138298034668,
22
+ "learning_rate": 6.882352941176471e-06,
23
+ "loss": 0.7025,
24
+ "step": 40
25
+ },
26
+ {
27
+ "epoch": 0.17647058823529413,
28
+ "grad_norm": 45.79734802246094,
29
+ "learning_rate": 1.0411764705882353e-05,
30
+ "loss": 0.7232,
31
+ "step": 60
32
+ },
33
+ {
34
+ "epoch": 0.23529411764705882,
35
+ "grad_norm": 80.78459167480469,
36
+ "learning_rate": 1.3941176470588236e-05,
37
+ "loss": 0.6618,
38
+ "step": 80
39
+ },
40
+ {
41
+ "epoch": 0.29411764705882354,
42
+ "grad_norm": 78.52603149414062,
43
+ "learning_rate": 1.747058823529412e-05,
44
+ "loss": 0.6135,
45
+ "step": 100
46
+ },
47
+ {
48
+ "epoch": 0.35294117647058826,
49
+ "grad_norm": 23.83449363708496,
50
+ "learning_rate": 2.1e-05,
51
+ "loss": 0.7209,
52
+ "step": 120
53
+ },
54
+ {
55
+ "epoch": 0.4117647058823529,
56
+ "grad_norm": 83.61427307128906,
57
+ "learning_rate": 2.452941176470588e-05,
58
+ "loss": 0.6285,
59
+ "step": 140
60
+ },
61
+ {
62
+ "epoch": 0.47058823529411764,
63
+ "grad_norm": 13.506064414978027,
64
+ "learning_rate": 2.8058823529411766e-05,
65
+ "loss": 0.6619,
66
+ "step": 160
67
+ },
68
+ {
69
+ "epoch": 0.5294117647058824,
70
+ "grad_norm": 7.046374320983887,
71
+ "learning_rate": 2.9999425303953133e-05,
72
+ "loss": 0.6799,
73
+ "step": 180
74
+ },
75
+ {
76
+ "epoch": 0.5882352941176471,
77
+ "grad_norm": 9.024923324584961,
78
+ "learning_rate": 2.999403345162272e-05,
79
+ "loss": 0.5675,
80
+ "step": 200
81
+ },
82
+ {
83
+ "epoch": 0.6470588235294118,
84
+ "grad_norm": 6.643726348876953,
85
+ "learning_rate": 2.9982967989381395e-05,
86
+ "loss": 0.6948,
87
+ "step": 220
88
+ },
89
+ {
90
+ "epoch": 0.7058823529411765,
91
+ "grad_norm": 4.687771797180176,
92
+ "learning_rate": 2.99662331043024e-05,
93
+ "loss": 0.5777,
94
+ "step": 240
95
+ },
96
+ {
97
+ "epoch": 0.7647058823529411,
98
+ "grad_norm": 8.102996826171875,
99
+ "learning_rate": 2.994383512871854e-05,
100
+ "loss": 0.6405,
101
+ "step": 260
102
+ },
103
+ {
104
+ "epoch": 0.8235294117647058,
105
+ "grad_norm": 4.696077823638916,
106
+ "learning_rate": 2.9915782537826074e-05,
107
+ "loss": 0.624,
108
+ "step": 280
109
+ },
110
+ {
111
+ "epoch": 0.8823529411764706,
112
+ "grad_norm": 11.583894729614258,
113
+ "learning_rate": 2.9882085946477782e-05,
114
+ "loss": 0.5285,
115
+ "step": 300
116
+ },
117
+ {
118
+ "epoch": 0.9411764705882353,
119
+ "grad_norm": 5.211498737335205,
120
+ "learning_rate": 2.9842758105166408e-05,
121
+ "loss": 0.6815,
122
+ "step": 320
123
+ },
124
+ {
125
+ "epoch": 1.0,
126
+ "grad_norm": 7.949342250823975,
127
+ "learning_rate": 2.979781389519997e-05,
128
+ "loss": 0.5037,
129
+ "step": 340
130
+ },
131
+ {
132
+ "epoch": 1.0,
133
+ "eval_category_set_accuracy": 0.10927152317880795,
134
+ "eval_is_valid_accuracy": 0.7814569536423841,
135
+ "eval_loss": 0.35260045528411865,
136
+ "eval_macro_f1": 0.18852975092832924,
137
+ "eval_micro_f1": 0.24468085106382978,
138
+ "eval_runtime": 6.535,
139
+ "eval_samples_per_second": 92.425,
140
+ "eval_steps_per_second": 11.63,
141
+ "step": 340
142
+ },
143
+ {
144
+ "epoch": 1.0588235294117647,
145
+ "grad_norm": 7.074536323547363,
146
+ "learning_rate": 2.974727032307081e-05,
147
+ "loss": 0.7041,
148
+ "step": 360
149
+ },
150
+ {
151
+ "epoch": 1.1176470588235294,
152
+ "grad_norm": 5.967779636383057,
153
+ "learning_rate": 2.969114651402049e-05,
154
+ "loss": 0.5153,
155
+ "step": 380
156
+ },
157
+ {
158
+ "epoch": 1.1764705882352942,
159
+ "grad_norm": 6.250365257263184,
160
+ "learning_rate": 2.962946370480297e-05,
161
+ "loss": 0.5673,
162
+ "step": 400
163
+ },
164
+ {
165
+ "epoch": 1.2352941176470589,
166
+ "grad_norm": 6.270878791809082,
167
+ "learning_rate": 2.9562245235648804e-05,
168
+ "loss": 0.5831,
169
+ "step": 420
170
+ },
171
+ {
172
+ "epoch": 1.2941176470588236,
173
+ "grad_norm": 7.3866682052612305,
174
+ "learning_rate": 2.948951654143341e-05,
175
+ "loss": 0.4377,
176
+ "step": 440
177
+ },
178
+ {
179
+ "epoch": 1.3529411764705883,
180
+ "grad_norm": 8.181004524230957,
181
+ "learning_rate": 2.9411305142052725e-05,
182
+ "loss": 0.6458,
183
+ "step": 460
184
+ },
185
+ {
186
+ "epoch": 1.4117647058823528,
187
+ "grad_norm": 12.09321117401123,
188
+ "learning_rate": 2.9327640632009927e-05,
189
+ "loss": 0.476,
190
+ "step": 480
191
+ },
192
+ {
193
+ "epoch": 1.4705882352941178,
194
+ "grad_norm": 5.451741695404053,
195
+ "learning_rate": 2.9238554669217142e-05,
196
+ "loss": 0.531,
197
+ "step": 500
198
+ },
199
+ {
200
+ "epoch": 1.5294117647058822,
201
+ "grad_norm": 8.42138385772705,
202
+ "learning_rate": 2.9144080963016352e-05,
203
+ "loss": 0.5418,
204
+ "step": 520
205
+ },
206
+ {
207
+ "epoch": 1.5882352941176472,
208
+ "grad_norm": 6.258157730102539,
209
+ "learning_rate": 2.904425526142411e-05,
210
+ "loss": 0.3755,
211
+ "step": 540
212
+ },
213
+ {
214
+ "epoch": 1.6470588235294117,
215
+ "grad_norm": 3.9103658199310303,
216
+ "learning_rate": 2.8939115337604764e-05,
217
+ "loss": 0.681,
218
+ "step": 560
219
+ },
220
+ {
221
+ "epoch": 1.7058823529411766,
222
+ "grad_norm": 8.98405647277832,
223
+ "learning_rate": 2.882870097557745e-05,
224
+ "loss": 0.4283,
225
+ "step": 580
226
+ },
227
+ {
228
+ "epoch": 1.7647058823529411,
229
+ "grad_norm": 7.521501541137695,
230
+ "learning_rate": 2.8713053955162174e-05,
231
+ "loss": 0.5017,
232
+ "step": 600
233
+ },
234
+ {
235
+ "epoch": 1.8235294117647058,
236
+ "grad_norm": 12.84186840057373,
237
+ "learning_rate": 2.859221803617068e-05,
238
+ "loss": 0.4444,
239
+ "step": 620
240
+ },
241
+ {
242
+ "epoch": 1.8823529411764706,
243
+ "grad_norm": 13.120209693908691,
244
+ "learning_rate": 2.8466238941848144e-05,
245
+ "loss": 0.3279,
246
+ "step": 640
247
+ },
248
+ {
249
+ "epoch": 1.9411764705882353,
250
+ "grad_norm": 13.905231475830078,
251
+ "learning_rate": 2.833516434157188e-05,
252
+ "loss": 0.543,
253
+ "step": 660
254
+ },
255
+ {
256
+ "epoch": 2.0,
257
+ "grad_norm": 13.589735984802246,
258
+ "learning_rate": 2.819904383281368e-05,
259
+ "loss": 0.2781,
260
+ "step": 680
261
+ },
262
+ {
263
+ "epoch": 2.0,
264
+ "eval_category_set_accuracy": 0.3741721854304636,
265
+ "eval_is_valid_accuracy": 0.8493377483443708,
266
+ "eval_loss": 0.29620495438575745,
267
+ "eval_macro_f1": 0.5247712467274231,
268
+ "eval_micro_f1": 0.5559729526671675,
269
+ "eval_runtime": 6.4845,
270
+ "eval_samples_per_second": 93.145,
271
+ "eval_steps_per_second": 11.72,
272
+ "step": 680
273
+ },
274
+ {
275
+ "epoch": 2.0588235294117645,
276
+ "grad_norm": 5.380058765411377,
277
+ "learning_rate": 2.805792892237257e-05,
278
+ "loss": 0.5126,
279
+ "step": 700
280
+ },
281
+ {
282
+ "epoch": 2.1176470588235294,
283
+ "grad_norm": 13.700803756713867,
284
+ "learning_rate": 2.7911873006885093e-05,
285
+ "loss": 0.2966,
286
+ "step": 720
287
+ },
288
+ {
289
+ "epoch": 2.176470588235294,
290
+ "grad_norm": 8.947591781616211,
291
+ "learning_rate": 2.776093135262048e-05,
292
+ "loss": 0.4139,
293
+ "step": 740
294
+ },
295
+ {
296
+ "epoch": 2.235294117647059,
297
+ "grad_norm": 5.173595428466797,
298
+ "learning_rate": 2.760516107456839e-05,
299
+ "loss": 0.3338,
300
+ "step": 760
301
+ },
302
+ {
303
+ "epoch": 2.2941176470588234,
304
+ "grad_norm": 13.049681663513184,
305
+ "learning_rate": 2.7444621114827112e-05,
306
+ "loss": 0.1979,
307
+ "step": 780
308
+ },
309
+ {
310
+ "epoch": 2.3529411764705883,
311
+ "grad_norm": 5.6180195808410645,
312
+ "learning_rate": 2.727937222030039e-05,
313
+ "loss": 0.4683,
314
+ "step": 800
315
+ },
316
+ {
317
+ "epoch": 2.411764705882353,
318
+ "grad_norm": 11.237857818603516,
319
+ "learning_rate": 2.7109476919711298e-05,
320
+ "loss": 0.222,
321
+ "step": 820
322
+ },
323
+ {
324
+ "epoch": 2.4705882352941178,
325
+ "grad_norm": 15.772666931152344,
326
+ "learning_rate": 2.6934999499941986e-05,
327
+ "loss": 0.3023,
328
+ "step": 840
329
+ },
330
+ {
331
+ "epoch": 2.5294117647058822,
332
+ "grad_norm": 9.203362464904785,
333
+ "learning_rate": 2.675600598170804e-05,
334
+ "loss": 0.3068,
335
+ "step": 860
336
+ },
337
+ {
338
+ "epoch": 2.588235294117647,
339
+ "grad_norm": 9.195798873901367,
340
+ "learning_rate": 2.657256409457684e-05,
341
+ "loss": 0.1588,
342
+ "step": 880
343
+ },
344
+ {
345
+ "epoch": 2.6470588235294117,
346
+ "grad_norm": 13.020779609680176,
347
+ "learning_rate": 2.6384743251339303e-05,
348
+ "loss": 0.3574,
349
+ "step": 900
350
+ },
351
+ {
352
+ "epoch": 2.7058823529411766,
353
+ "grad_norm": 15.945825576782227,
354
+ "learning_rate": 2.619261452174468e-05,
355
+ "loss": 0.1668,
356
+ "step": 920
357
+ },
358
+ {
359
+ "epoch": 2.764705882352941,
360
+ "grad_norm": 9.987348556518555,
361
+ "learning_rate": 2.5996250605608395e-05,
362
+ "loss": 0.2546,
363
+ "step": 940
364
+ },
365
+ {
366
+ "epoch": 2.8235294117647056,
367
+ "grad_norm": 14.392949104309082,
368
+ "learning_rate": 2.5795725805303072e-05,
369
+ "loss": 0.2341,
370
+ "step": 960
371
+ },
372
+ {
373
+ "epoch": 2.8823529411764706,
374
+ "grad_norm": 51.86268997192383,
375
+ "learning_rate": 2.5591115997643192e-05,
376
+ "loss": 0.1137,
377
+ "step": 980
378
+ },
379
+ {
380
+ "epoch": 2.9411764705882355,
381
+ "grad_norm": 19.037843704223633,
382
+ "learning_rate": 2.5382498605173945e-05,
383
+ "loss": 0.2876,
384
+ "step": 1000
385
+ },
386
+ {
387
+ "epoch": 3.0,
388
+ "grad_norm": 72.03766632080078,
389
+ "learning_rate": 2.5169952566875248e-05,
390
+ "loss": 0.1107,
391
+ "step": 1020
392
+ },
393
+ {
394
+ "epoch": 3.0,
395
+ "eval_category_set_accuracy": 0.5463576158940397,
396
+ "eval_is_valid_accuracy": 0.9420529801324503,
397
+ "eval_loss": 0.18532413244247437,
398
+ "eval_macro_f1": 0.7386485042671416,
399
+ "eval_micro_f1": 0.7245033112582782,
400
+ "eval_runtime": 6.545,
401
+ "eval_samples_per_second": 92.284,
402
+ "eval_steps_per_second": 11.612,
403
+ "step": 1020
404
+ },
405
+ {
406
+ "epoch": 3.0588235294117645,
407
+ "grad_norm": 17.806486129760742,
408
+ "learning_rate": 2.495355830829193e-05,
409
+ "loss": 0.2421,
410
+ "step": 1040
411
+ },
412
+ {
413
+ "epoch": 3.1176470588235294,
414
+ "grad_norm": 7.106595993041992,
415
+ "learning_rate": 2.4733397711101392e-05,
416
+ "loss": 0.1021,
417
+ "step": 1060
418
+ },
419
+ {
420
+ "epoch": 3.176470588235294,
421
+ "grad_norm": 6.141881942749023,
422
+ "learning_rate": 2.450955408213033e-05,
423
+ "loss": 0.1925,
424
+ "step": 1080
425
+ },
426
+ {
427
+ "epoch": 3.235294117647059,
428
+ "grad_norm": 10.608316421508789,
429
+ "learning_rate": 2.4282112121832163e-05,
430
+ "loss": 0.1307,
431
+ "step": 1100
432
+ },
433
+ {
434
+ "epoch": 3.2941176470588234,
435
+ "grad_norm": 10.223675727844238,
436
+ "learning_rate": 2.4051157892237114e-05,
437
+ "loss": 0.0648,
438
+ "step": 1120
439
+ },
440
+ {
441
+ "epoch": 3.3529411764705883,
442
+ "grad_norm": 9.1436767578125,
443
+ "learning_rate": 2.3816778784387097e-05,
444
+ "loss": 0.1929,
445
+ "step": 1140
446
+ },
447
+ {
448
+ "epoch": 3.411764705882353,
449
+ "grad_norm": 10.77053451538086,
450
+ "learning_rate": 2.3579063485267752e-05,
451
+ "loss": 0.0614,
452
+ "step": 1160
453
+ },
454
+ {
455
+ "epoch": 3.4705882352941178,
456
+ "grad_norm": 20.1484432220459,
457
+ "learning_rate": 2.333810194425003e-05,
458
+ "loss": 0.162,
459
+ "step": 1180
460
+ },
461
+ {
462
+ "epoch": 3.5294117647058822,
463
+ "grad_norm": 14.523782730102539,
464
+ "learning_rate": 2.3093985339054165e-05,
465
+ "loss": 0.0994,
466
+ "step": 1200
467
+ },
468
+ {
469
+ "epoch": 3.588235294117647,
470
+ "grad_norm": 6.1038408279418945,
471
+ "learning_rate": 2.284680604124883e-05,
472
+ "loss": 0.071,
473
+ "step": 1220
474
+ },
475
+ {
476
+ "epoch": 3.6470588235294117,
477
+ "grad_norm": 3.1415257453918457,
478
+ "learning_rate": 2.259665758129855e-05,
479
+ "loss": 0.2435,
480
+ "step": 1240
481
+ },
482
+ {
483
+ "epoch": 3.7058823529411766,
484
+ "grad_norm": 15.256193161010742,
485
+ "learning_rate": 2.2343634613172596e-05,
486
+ "loss": 0.0711,
487
+ "step": 1260
488
+ },
489
+ {
490
+ "epoch": 3.764705882352941,
491
+ "grad_norm": 10.315400123596191,
492
+ "learning_rate": 2.208783287852876e-05,
493
+ "loss": 0.1267,
494
+ "step": 1280
495
+ },
496
+ {
497
+ "epoch": 3.8235294117647056,
498
+ "grad_norm": 16.806316375732422,
499
+ "learning_rate": 2.1829349170485545e-05,
500
+ "loss": 0.0968,
501
+ "step": 1300
502
+ },
503
+ {
504
+ "epoch": 3.8823529411764706,
505
+ "grad_norm": 1.3469294309616089,
506
+ "learning_rate": 2.156828129699655e-05,
507
+ "loss": 0.031,
508
+ "step": 1320
509
+ },
510
+ {
511
+ "epoch": 3.9411764705882355,
512
+ "grad_norm": 3.2037153244018555,
513
+ "learning_rate": 2.1304728043840734e-05,
514
+ "loss": 0.1532,
515
+ "step": 1340
516
+ },
517
+ {
518
+ "epoch": 4.0,
519
+ "grad_norm": 157.6475372314453,
520
+ "learning_rate": 2.1038789137242793e-05,
521
+ "loss": 0.0684,
522
+ "step": 1360
523
+ },
524
+ {
525
+ "epoch": 4.0,
526
+ "eval_category_set_accuracy": 0.695364238410596,
527
+ "eval_is_valid_accuracy": 0.9801324503311258,
528
+ "eval_loss": 0.16694329679012299,
529
+ "eval_macro_f1": 0.8161149900922228,
530
+ "eval_micro_f1": 0.8069241011984021,
531
+ "eval_runtime": 6.4686,
532
+ "eval_samples_per_second": 93.374,
533
+ "eval_steps_per_second": 11.749,
534
+ "step": 1360
535
+ },
536
+ {
537
+ "epoch": 4.0588235294117645,
538
+ "grad_norm": 5.109766960144043,
539
+ "learning_rate": 2.0770565206137636e-05,
540
+ "loss": 0.083,
541
+ "step": 1380
542
+ },
543
+ {
544
+ "epoch": 4.117647058823529,
545
+ "grad_norm": 2.3519508838653564,
546
+ "learning_rate": 2.0500157744093284e-05,
547
+ "loss": 0.0351,
548
+ "step": 1400
549
+ },
550
+ {
551
+ "epoch": 4.176470588235294,
552
+ "grad_norm": 8.30711555480957,
553
+ "learning_rate": 2.0227669070906566e-05,
554
+ "loss": 0.0666,
555
+ "step": 1420
556
+ },
557
+ {
558
+ "epoch": 4.235294117647059,
559
+ "grad_norm": 3.5932908058166504,
560
+ "learning_rate": 1.9953202293886253e-05,
561
+ "loss": 0.0736,
562
+ "step": 1440
563
+ },
564
+ {
565
+ "epoch": 4.294117647058823,
566
+ "grad_norm": 0.6442999839782715,
567
+ "learning_rate": 1.967686126883815e-05,
568
+ "loss": 0.0073,
569
+ "step": 1460
570
+ },
571
+ {
572
+ "epoch": 4.352941176470588,
573
+ "grad_norm": 37.24504470825195,
574
+ "learning_rate": 1.9398750560766973e-05,
575
+ "loss": 0.0795,
576
+ "step": 1480
577
+ },
578
+ {
579
+ "epoch": 4.411764705882353,
580
+ "grad_norm": 4.601501941680908,
581
+ "learning_rate": 1.911897540430986e-05,
582
+ "loss": 0.0181,
583
+ "step": 1500
584
+ },
585
+ {
586
+ "epoch": 4.470588235294118,
587
+ "grad_norm": 28.21880531311035,
588
+ "learning_rate": 1.8837641663916536e-05,
589
+ "loss": 0.0572,
590
+ "step": 1520
591
+ },
592
+ {
593
+ "epoch": 4.529411764705882,
594
+ "grad_norm": 1.3415755033493042,
595
+ "learning_rate": 1.855485579379113e-05,
596
+ "loss": 0.0844,
597
+ "step": 1540
598
+ },
599
+ {
600
+ "epoch": 4.588235294117647,
601
+ "grad_norm": 1.8818895816802979,
602
+ "learning_rate": 1.8270724797610845e-05,
603
+ "loss": 0.0281,
604
+ "step": 1560
605
+ },
606
+ {
607
+ "epoch": 4.647058823529412,
608
+ "grad_norm": 5.257987022399902,
609
+ "learning_rate": 1.7985356188036712e-05,
610
+ "loss": 0.0822,
611
+ "step": 1580
612
+ },
613
+ {
614
+ "epoch": 4.705882352941177,
615
+ "grad_norm": 0.3357090651988983,
616
+ "learning_rate": 1.769885794603174e-05,
617
+ "loss": 0.057,
618
+ "step": 1600
619
+ },
620
+ {
621
+ "epoch": 4.764705882352941,
622
+ "grad_norm": 5.4116106033325195,
623
+ "learning_rate": 1.741133848000188e-05,
624
+ "loss": 0.0724,
625
+ "step": 1620
626
+ },
627
+ {
628
+ "epoch": 4.823529411764706,
629
+ "grad_norm": 35.08403396606445,
630
+ "learning_rate": 1.7122906584775248e-05,
631
+ "loss": 0.0353,
632
+ "step": 1640
633
+ },
634
+ {
635
+ "epoch": 4.882352941176471,
636
+ "grad_norm": 0.4787577688694,
637
+ "learning_rate": 1.6833671400435095e-05,
638
+ "loss": 0.0117,
639
+ "step": 1660
640
+ },
641
+ {
642
+ "epoch": 4.9411764705882355,
643
+ "grad_norm": 5.700762748718262,
644
+ "learning_rate": 1.6543742371022217e-05,
645
+ "loss": 0.0454,
646
+ "step": 1680
647
+ },
648
+ {
649
+ "epoch": 5.0,
650
+ "grad_norm": 0.41345372796058655,
651
+ "learning_rate": 1.625322920312222e-05,
652
+ "loss": 0.0162,
653
+ "step": 1700
654
+ },
655
+ {
656
+ "epoch": 5.0,
657
+ "eval_category_set_accuracy": 0.7831125827814569,
658
+ "eval_is_valid_accuracy": 0.9668874172185431,
659
+ "eval_loss": 0.35251227021217346,
660
+ "eval_macro_f1": 0.8575976396227553,
661
+ "eval_micro_f1": 0.8575518969219756,
662
+ "eval_runtime": 6.5918,
663
+ "eval_samples_per_second": 91.629,
664
+ "eval_steps_per_second": 11.529,
665
+ "step": 1700
666
+ },
667
+ {
668
+ "epoch": 5.0588235294117645,
669
+ "grad_norm": 4.610188961029053,
670
+ "learning_rate": 1.596224182435359e-05,
671
+ "loss": 0.0398,
672
+ "step": 1720
673
+ },
674
+ {
675
+ "epoch": 5.117647058823529,
676
+ "grad_norm": 0.016779739409685135,
677
+ "learning_rate": 1.5670890341771996e-05,
678
+ "loss": 0.0107,
679
+ "step": 1740
680
+ },
681
+ {
682
+ "epoch": 5.176470588235294,
683
+ "grad_norm": 26.066761016845703,
684
+ "learning_rate": 1.53792850002068e-05,
685
+ "loss": 0.0339,
686
+ "step": 1760
687
+ },
688
+ {
689
+ "epoch": 5.235294117647059,
690
+ "grad_norm": 0.02171199768781662,
691
+ "learning_rate": 1.508753614054534e-05,
692
+ "loss": 0.0474,
693
+ "step": 1780
694
+ },
695
+ {
696
+ "epoch": 5.294117647058823,
697
+ "grad_norm": 2.298548936843872,
698
+ "learning_rate": 1.4795754157980954e-05,
699
+ "loss": 0.0084,
700
+ "step": 1800
701
+ },
702
+ {
703
+ "epoch": 5.352941176470588,
704
+ "grad_norm": 2.896538019180298,
705
+ "learning_rate": 1.4504049460240376e-05,
706
+ "loss": 0.0416,
707
+ "step": 1820
708
+ },
709
+ {
710
+ "epoch": 5.411764705882353,
711
+ "grad_norm": 0.12819333374500275,
712
+ "learning_rate": 1.4212532425806462e-05,
713
+ "loss": 0.0292,
714
+ "step": 1840
715
+ },
716
+ {
717
+ "epoch": 5.470588235294118,
718
+ "grad_norm": 3.9555938243865967,
719
+ "learning_rate": 1.3921313362151915e-05,
720
+ "loss": 0.032,
721
+ "step": 1860
722
+ },
723
+ {
724
+ "epoch": 5.529411764705882,
725
+ "grad_norm": 0.028258875012397766,
726
+ "learning_rate": 1.3630502463999932e-05,
727
+ "loss": 0.0234,
728
+ "step": 1880
729
+ },
730
+ {
731
+ "epoch": 5.588235294117647,
732
+ "grad_norm": 0.018097462132573128,
733
+ "learning_rate": 1.334020977162749e-05,
734
+ "loss": 0.0009,
735
+ "step": 1900
736
+ },
737
+ {
738
+ "epoch": 5.647058823529412,
739
+ "grad_norm": 0.038703106343746185,
740
+ "learning_rate": 1.3050545129227066e-05,
741
+ "loss": 0.0442,
742
+ "step": 1920
743
+ },
744
+ {
745
+ "epoch": 5.705882352941177,
746
+ "grad_norm": 0.0021211367566138506,
747
+ "learning_rate": 1.2761618143342567e-05,
748
+ "loss": 0.012,
749
+ "step": 1940
750
+ },
751
+ {
752
+ "epoch": 5.764705882352941,
753
+ "grad_norm": 3.898015260696411,
754
+ "learning_rate": 1.2473538141395206e-05,
755
+ "loss": 0.0339,
756
+ "step": 1960
757
+ },
758
+ {
759
+ "epoch": 5.823529411764706,
760
+ "grad_norm": 0.01642230711877346,
761
+ "learning_rate": 1.2186414130314924e-05,
762
+ "loss": 0.0087,
763
+ "step": 1980
764
+ },
765
+ {
766
+ "epoch": 5.882352941176471,
767
+ "grad_norm": 0.027277540415525436,
768
+ "learning_rate": 1.1900354755293178e-05,
769
+ "loss": 0.0007,
770
+ "step": 2000
771
+ },
772
+ {
773
+ "epoch": 5.9411764705882355,
774
+ "grad_norm": 0.20538949966430664,
775
+ "learning_rate": 1.1615468258672527e-05,
776
+ "loss": 0.0312,
777
+ "step": 2020
778
+ },
779
+ {
780
+ "epoch": 6.0,
781
+ "grad_norm": 0.0008514047949574888,
782
+ "learning_rate": 1.133186243898869e-05,
783
+ "loss": 0.0286,
784
+ "step": 2040
785
+ },
786
+ {
787
+ "epoch": 6.0,
788
+ "eval_category_set_accuracy": 0.7781456953642384,
789
+ "eval_is_valid_accuracy": 0.9768211920529801,
790
+ "eval_loss": 0.29321521520614624,
791
+ "eval_macro_f1": 0.8661252418309986,
792
+ "eval_micro_f1": 0.8630705394190872,
793
+ "eval_runtime": 6.5015,
794
+ "eval_samples_per_second": 92.902,
795
+ "eval_steps_per_second": 11.69,
796
+ "step": 2040
797
+ },
798
+ {
799
+ "epoch": 6.0588235294117645,
800
+ "grad_norm": 2.5546793937683105,
801
+ "learning_rate": 1.1049644610180553e-05,
802
+ "loss": 0.0213,
803
+ "step": 2060
804
+ },
805
+ {
806
+ "epoch": 6.117647058823529,
807
+ "grad_norm": 0.0012535699643194675,
808
+ "learning_rate": 1.076892156098349e-05,
809
+ "loss": 0.0008,
810
+ "step": 2080
811
+ },
812
+ {
813
+ "epoch": 6.176470588235294,
814
+ "grad_norm": 2.8291475772857666,
815
+ "learning_rate": 1.0489799514521475e-05,
816
+ "loss": 0.0072,
817
+ "step": 2100
818
+ },
819
+ {
820
+ "epoch": 6.235294117647059,
821
+ "grad_norm": 0.28834396600723267,
822
+ "learning_rate": 1.0212384088113177e-05,
823
+ "loss": 0.0012,
824
+ "step": 2120
825
+ },
826
+ {
827
+ "epoch": 6.294117647058823,
828
+ "grad_norm": 0.00216933386400342,
829
+ "learning_rate": 9.936780253307296e-06,
830
+ "loss": 0.0001,
831
+ "step": 2140
832
+ },
833
+ {
834
+ "epoch": 6.352941176470588,
835
+ "grad_norm": 0.6339454650878906,
836
+ "learning_rate": 9.663092296162252e-06,
837
+ "loss": 0.0204,
838
+ "step": 2160
839
+ },
840
+ {
841
+ "epoch": 6.411764705882353,
842
+ "grad_norm": 0.0003532178234308958,
843
+ "learning_rate": 9.391423777785298e-06,
844
+ "loss": 0.0096,
845
+ "step": 2180
846
+ },
847
+ {
848
+ "epoch": 6.470588235294118,
849
+ "grad_norm": 0.05404308810830116,
850
+ "learning_rate": 9.12187749514585e-06,
851
+ "loss": 0.0053,
852
+ "step": 2200
853
+ },
854
+ {
855
+ "epoch": 6.529411764705882,
856
+ "grad_norm": 0.00025081689818762243,
857
+ "learning_rate": 8.85455544217807e-06,
858
+ "loss": 0.0085,
859
+ "step": 2220
860
+ },
861
+ {
862
+ "epoch": 6.588235294117647,
863
+ "grad_norm": 8.275144500657916e-05,
864
+ "learning_rate": 8.589558771187245e-06,
865
+ "loss": 0.0093,
866
+ "step": 2240
867
+ },
868
+ {
869
+ "epoch": 6.647058823529412,
870
+ "grad_norm": 0.014114965684711933,
871
+ "learning_rate": 8.326987754574632e-06,
872
+ "loss": 0.0119,
873
+ "step": 2260
874
+ },
875
+ {
876
+ "epoch": 6.705882352941177,
877
+ "grad_norm": 0.003770261537283659,
878
+ "learning_rate": 8.066941746895304e-06,
879
+ "loss": 0.0094,
880
+ "step": 2280
881
+ },
882
+ {
883
+ "epoch": 6.764705882352941,
884
+ "grad_norm": 2.6549384593963623,
885
+ "learning_rate": 7.809519147263228e-06,
886
+ "loss": 0.0124,
887
+ "step": 2300
888
+ },
889
+ {
890
+ "epoch": 6.823529411764706,
891
+ "grad_norm": 0.10680641978979111,
892
+ "learning_rate": 7.554817362117915e-06,
893
+ "loss": 0.0186,
894
+ "step": 2320
895
+ },
896
+ {
897
+ "epoch": 6.882352941176471,
898
+ "grad_norm": 0.00080859538866207,
899
+ "learning_rate": 7.302932768366711e-06,
900
+ "loss": 0.0011,
901
+ "step": 2340
902
+ },
903
+ {
904
+ "epoch": 6.9411764705882355,
905
+ "grad_norm": 41.11537551879883,
906
+ "learning_rate": 7.05396067691659e-06,
907
+ "loss": 0.0202,
908
+ "step": 2360
909
+ },
910
+ {
911
+ "epoch": 7.0,
912
+ "grad_norm": 0.0007114345789887011,
913
+ "learning_rate": 6.807995296609361e-06,
914
+ "loss": 0.0074,
915
+ "step": 2380
916
+ },
917
+ {
918
+ "epoch": 7.0,
919
+ "eval_category_set_accuracy": 0.8162251655629139,
920
+ "eval_is_valid_accuracy": 0.9768211920529801,
921
+ "eval_loss": 0.30258405208587646,
922
+ "eval_macro_f1": 0.8929467566472393,
923
+ "eval_micro_f1": 0.8920254057868737,
924
+ "eval_runtime": 6.6165,
925
+ "eval_samples_per_second": 91.287,
926
+ "eval_steps_per_second": 11.486,
927
+ "step": 2380
928
+ },
929
+ {
930
+ "epoch": 7.0588235294117645,
931
+ "grad_norm": 0.011070066131651402,
932
+ "learning_rate": 6.565129698573884e-06,
933
+ "loss": 0.0022,
934
+ "step": 2400
935
+ },
936
+ {
937
+ "epoch": 7.117647058823529,
938
+ "grad_norm": 0.003187597496435046,
939
+ "learning_rate": 6.325455781008763e-06,
940
+ "loss": 0.0007,
941
+ "step": 2420
942
+ },
943
+ {
944
+ "epoch": 7.176470588235294,
945
+ "grad_norm": 25.223493576049805,
946
+ "learning_rate": 6.0890642344088776e-06,
947
+ "loss": 0.0035,
948
+ "step": 2440
949
+ },
950
+ {
951
+ "epoch": 7.235294117647059,
952
+ "grad_norm": 0.015322019346058369,
953
+ "learning_rate": 5.8560445072489295e-06,
954
+ "loss": 0.0034,
955
+ "step": 2460
956
+ },
957
+ {
958
+ "epoch": 7.294117647058823,
959
+ "grad_norm": 2.8156142434454523e-05,
960
+ "learning_rate": 5.626484772136881e-06,
961
+ "loss": 0.0018,
962
+ "step": 2480
963
+ },
964
+ {
965
+ "epoch": 7.352941176470588,
966
+ "grad_norm": 0.03186884522438049,
967
+ "learning_rate": 5.400471892450251e-06,
968
+ "loss": 0.0011,
969
+ "step": 2500
970
+ },
971
+ {
972
+ "epoch": 7.411764705882353,
973
+ "grad_norm": 2.5159380356853944e-07,
974
+ "learning_rate": 5.1780913894677846e-06,
975
+ "loss": 0.0001,
976
+ "step": 2520
977
+ },
978
+ {
979
+ "epoch": 7.470588235294118,
980
+ "grad_norm": 0.18598201870918274,
981
+ "learning_rate": 4.959427410008931e-06,
982
+ "loss": 0.0008,
983
+ "step": 2540
984
+ },
985
+ {
986
+ "epoch": 7.529411764705882,
987
+ "grad_norm": 1.9272384643554688,
988
+ "learning_rate": 4.7445626945934404e-06,
989
+ "loss": 0.0001,
990
+ "step": 2560
991
+ },
992
+ {
993
+ "epoch": 7.588235294117647,
994
+ "grad_norm": 7.508861017413437e-05,
995
+ "learning_rate": 4.5335785461330934e-06,
996
+ "loss": 0.0,
997
+ "step": 2580
998
+ },
999
+ {
1000
+ "epoch": 7.647058823529412,
1001
+ "grad_norm": 0.0002854722843039781,
1002
+ "learning_rate": 4.326554799167376e-06,
1003
+ "loss": 0.0064,
1004
+ "step": 2600
1005
+ },
1006
+ {
1007
+ "epoch": 7.705882352941177,
1008
+ "grad_norm": 0.0007522217347286642,
1009
+ "learning_rate": 4.123569789654823e-06,
1010
+ "loss": 0.0001,
1011
+ "step": 2620
1012
+ },
1013
+ {
1014
+ "epoch": 7.764705882352941,
1015
+ "grad_norm": 0.00011735141015378758,
1016
+ "learning_rate": 3.924700325331349e-06,
1017
+ "loss": 0.001,
1018
+ "step": 2640
1019
+ },
1020
+ {
1021
+ "epoch": 7.823529411764706,
1022
+ "grad_norm": 2.814086474245414e-05,
1023
+ "learning_rate": 3.7300216566469e-06,
1024
+ "loss": 0.0092,
1025
+ "step": 2660
1026
+ },
1027
+ {
1028
+ "epoch": 7.882352941176471,
1029
+ "grad_norm": 4.8900161345954984e-05,
1030
+ "learning_rate": 3.539607448291349e-06,
1031
+ "loss": 0.0,
1032
+ "step": 2680
1033
+ },
1034
+ {
1035
+ "epoch": 7.9411764705882355,
1036
+ "grad_norm": 0.007463980466127396,
1037
+ "learning_rate": 3.3535297513204162e-06,
1038
+ "loss": 0.0011,
1039
+ "step": 2700
1040
+ },
1041
+ {
1042
+ "epoch": 8.0,
1043
+ "grad_norm": 5.794290700578131e-07,
1044
+ "learning_rate": 3.171858975892189e-06,
1045
+ "loss": 0.0,
1046
+ "step": 2720
1047
+ },
1048
+ {
1049
+ "epoch": 8.0,
1050
+ "eval_category_set_accuracy": 0.8228476821192053,
1051
+ "eval_is_valid_accuracy": 0.9718543046357616,
1052
+ "eval_loss": 0.37849172949790955,
1053
+ "eval_macro_f1": 0.8988780276395113,
1054
+ "eval_micro_f1": 0.8986340762041697,
1055
+ "eval_runtime": 6.5574,
1056
+ "eval_samples_per_second": 92.109,
1057
+ "eval_steps_per_second": 11.59,
1058
+ "step": 2720
1059
+ },
1060
+ {
1061
+ "epoch": 8.058823529411764,
1062
+ "grad_norm": 0.00019079445337411016,
1063
+ "learning_rate": 2.9946638646245446e-06,
1064
+ "loss": 0.0001,
1065
+ "step": 2740
1066
+ },
1067
+ {
1068
+ "epoch": 8.117647058823529,
1069
+ "grad_norm": 0.0003336981462780386,
1070
+ "learning_rate": 2.8220114665835465e-06,
1071
+ "loss": 0.0,
1072
+ "step": 2760
1073
+ },
1074
+ {
1075
+ "epoch": 8.176470588235293,
1076
+ "grad_norm": 0.0020561586134135723,
1077
+ "learning_rate": 2.6539671119126676e-06,
1078
+ "loss": 0.0001,
1079
+ "step": 2780
1080
+ },
1081
+ {
1082
+ "epoch": 8.235294117647058,
1083
+ "grad_norm": 0.0002150319196516648,
1084
+ "learning_rate": 2.4905943871124517e-06,
1085
+ "loss": 0.0,
1086
+ "step": 2800
1087
+ },
1088
+ {
1089
+ "epoch": 8.294117647058824,
1090
+ "grad_norm": 0.000325396511470899,
1091
+ "learning_rate": 2.3319551109799187e-06,
1092
+ "loss": 0.0,
1093
+ "step": 2820
1094
+ },
1095
+ {
1096
+ "epoch": 8.352941176470589,
1097
+ "grad_norm": 0.0030735947657376528,
1098
+ "learning_rate": 2.1781093112169132e-06,
1099
+ "loss": 0.0,
1100
+ "step": 2840
1101
+ },
1102
+ {
1103
+ "epoch": 8.411764705882353,
1104
+ "grad_norm": 5.6676035455893725e-05,
1105
+ "learning_rate": 2.029115201716118e-06,
1106
+ "loss": 0.0,
1107
+ "step": 2860
1108
+ },
1109
+ {
1110
+ "epoch": 8.470588235294118,
1111
+ "grad_norm": 0.02090725675225258,
1112
+ "learning_rate": 1.8850291605334468e-06,
1113
+ "loss": 0.0,
1114
+ "step": 2880
1115
+ },
1116
+ {
1117
+ "epoch": 8.529411764705882,
1118
+ "grad_norm": 0.0006118013989180326,
1119
+ "learning_rate": 1.7459057085550838e-06,
1120
+ "loss": 0.0011,
1121
+ "step": 2900
1122
+ },
1123
+ {
1124
+ "epoch": 8.588235294117647,
1125
+ "grad_norm": 0.00014390128490049392,
1126
+ "learning_rate": 1.611797488867266e-06,
1127
+ "loss": 0.0003,
1128
+ "step": 2920
1129
+ },
1130
+ {
1131
+ "epoch": 8.647058823529411,
1132
+ "grad_norm": 0.045788347721099854,
1133
+ "learning_rate": 1.4827552468365902e-06,
1134
+ "loss": 0.0001,
1135
+ "step": 2940
1136
+ },
1137
+ {
1138
+ "epoch": 8.705882352941176,
1139
+ "grad_norm": 0.0006639071507379413,
1140
+ "learning_rate": 1.3588278109084385e-06,
1141
+ "loss": 0.0,
1142
+ "step": 2960
1143
+ },
1144
+ {
1145
+ "epoch": 8.764705882352942,
1146
+ "grad_norm": 0.001100758439861238,
1147
+ "learning_rate": 1.2400620741306939e-06,
1148
+ "loss": 0.0,
1149
+ "step": 2980
1150
+ },
1151
+ {
1152
+ "epoch": 8.823529411764707,
1153
+ "grad_norm": 0.0019803063478320837,
1154
+ "learning_rate": 1.1265029764098495e-06,
1155
+ "loss": 0.0,
1156
+ "step": 3000
1157
+ },
1158
+ {
1159
+ "epoch": 8.882352941176471,
1160
+ "grad_norm": 0.000128675572341308,
1161
+ "learning_rate": 1.0181934875061416e-06,
1162
+ "loss": 0.0,
1163
+ "step": 3020
1164
+ },
1165
+ {
1166
+ "epoch": 8.941176470588236,
1167
+ "grad_norm": 4.8775407776702195e-05,
1168
+ "learning_rate": 9.151745907741538e-07,
1169
+ "loss": 0.0,
1170
+ "step": 3040
1171
+ },
1172
+ {
1173
+ "epoch": 9.0,
1174
+ "grad_norm": 1.1025559842892108e-06,
1175
+ "learning_rate": 8.174852676550975e-07,
1176
+ "loss": 0.0,
1177
+ "step": 3060
1178
+ },
1179
+ {
1180
+ "epoch": 9.0,
1181
+ "eval_category_set_accuracy": 0.8261589403973509,
1182
+ "eval_is_valid_accuracy": 0.9735099337748344,
1183
+ "eval_loss": 0.39685797691345215,
1184
+ "eval_macro_f1": 0.9010823249433229,
1185
+ "eval_micro_f1": 0.9010830324909748,
1186
+ "eval_runtime": 6.4964,
1187
+ "eval_samples_per_second": 92.974,
1188
+ "eval_steps_per_second": 11.699,
1189
+ "step": 3060
1190
+ },
1191
+ {
1192
+ "epoch": 9.058823529411764,
1193
+ "grad_norm": 0.0011990389320999384,
1194
+ "learning_rate": 7.251624829265752e-07,
1195
+ "loss": 0.0,
1196
+ "step": 3080
1197
+ },
1198
+ {
1199
+ "epoch": 9.117647058823529,
1200
+ "grad_norm": 3.243568471589242e-06,
1201
+ "learning_rate": 6.382411707154245e-07,
1202
+ "loss": 0.0,
1203
+ "step": 3100
1204
+ },
1205
+ {
1206
+ "epoch": 9.176470588235293,
1207
+ "grad_norm": 0.015571022406220436,
1208
+ "learning_rate": 5.5675422127896e-07,
1209
+ "loss": 0.0,
1210
+ "step": 3120
1211
+ },
1212
+ {
1213
+ "epoch": 9.235294117647058,
1214
+ "grad_norm": 0.0005891515756957233,
1215
+ "learning_rate": 4.807324685595926e-07,
1216
+ "loss": 0.0,
1217
+ "step": 3140
1218
+ },
1219
+ {
1220
+ "epoch": 9.294117647058824,
1221
+ "grad_norm": 0.00015362814883701503,
1222
+ "learning_rate": 4.102046785175345e-07,
1223
+ "loss": 0.0,
1224
+ "step": 3160
1225
+ },
1226
+ {
1227
+ "epoch": 9.352941176470589,
1228
+ "grad_norm": 0.0006564821815118194,
1229
+ "learning_rate": 3.45197538246011e-07,
1230
+ "loss": 0.0,
1231
+ "step": 3180
1232
+ },
1233
+ {
1234
+ "epoch": 9.411764705882353,
1235
+ "grad_norm": 2.0837702322751284e-05,
1236
+ "learning_rate": 2.857356458731064e-07,
1237
+ "loss": 0.0,
1238
+ "step": 3200
1239
+ },
1240
+ {
1241
+ "epoch": 9.470588235294118,
1242
+ "grad_norm": 0.02459844760596752,
1243
+ "learning_rate": 2.318415012540448e-07,
1244
+ "loss": 0.0,
1245
+ "step": 3220
1246
+ },
1247
+ {
1248
+ "epoch": 9.529411764705882,
1249
+ "grad_norm": 0.0006321249529719353,
1250
+ "learning_rate": 1.8353549745744325e-07,
1251
+ "loss": 0.0,
1252
+ "step": 3240
1253
+ },
1254
+ {
1255
+ "epoch": 9.588235294117647,
1256
+ "grad_norm": 4.011229520983761e-06,
1257
+ "learning_rate": 1.408359130487613e-07,
1258
+ "loss": 0.0,
1259
+ "step": 3260
1260
+ },
1261
+ {
1262
+ "epoch": 9.647058823529411,
1263
+ "grad_norm": 0.012717085890471935,
1264
+ "learning_rate": 1.0375890517384023e-07,
1265
+ "loss": 0.0,
1266
+ "step": 3280
1267
+ },
1268
+ {
1269
+ "epoch": 9.705882352941176,
1270
+ "grad_norm": 2.548071734054247e-06,
1271
+ "learning_rate": 7.231850344519674e-08,
1272
+ "loss": 0.0,
1273
+ "step": 3300
1274
+ },
1275
+ {
1276
+ "epoch": 9.764705882352942,
1277
+ "grad_norm": 0.0015178547473624349,
1278
+ "learning_rate": 4.652660463333292e-08,
1279
+ "loss": 0.0,
1280
+ "step": 3320
1281
+ },
1282
+ {
1283
+ "epoch": 9.823529411764707,
1284
+ "grad_norm": 6.803183350712061e-05,
1285
+ "learning_rate": 2.639296816509873e-08,
1286
+ "loss": 0.0,
1287
+ "step": 3340
1288
+ },
1289
+ {
1290
+ "epoch": 9.882352941176471,
1291
+ "grad_norm": 0.001965134870260954,
1292
+ "learning_rate": 1.192521243082445e-08,
1293
+ "loss": 0.0,
1294
+ "step": 3360
1295
+ },
1296
+ {
1297
+ "epoch": 9.941176470588236,
1298
+ "grad_norm": 0.00012255716137588024,
1299
+ "learning_rate": 3.1288119015748794e-09,
1300
+ "loss": 0.0,
1301
+ "step": 3380
1302
+ },
1303
+ {
1304
+ "epoch": 10.0,
1305
+ "grad_norm": 1.7073415392587776e-06,
1306
+ "learning_rate": 7.09505767138019e-12,
1307
+ "loss": 0.0,
1308
+ "step": 3400
1309
+ },
1310
+ {
1311
+ "epoch": 10.0,
1312
+ "eval_category_set_accuracy": 0.8261589403973509,
1313
+ "eval_is_valid_accuracy": 0.9735099337748344,
1314
+ "eval_loss": 0.396047443151474,
1315
+ "eval_macro_f1": 0.9011001299947559,
1316
+ "eval_micro_f1": 0.9012256669069935,
1317
+ "eval_runtime": 6.474,
1318
+ "eval_samples_per_second": 93.296,
1319
+ "eval_steps_per_second": 11.739,
1320
+ "step": 3400
1321
+ }
1322
+ ],
1323
+ "logging_steps": 20,
1324
+ "max_steps": 3400,
1325
+ "num_input_tokens_seen": 0,
1326
+ "num_train_epochs": 10,
1327
+ "save_steps": 500,
1328
+ "stateful_callbacks": {
1329
+ "TrainerControl": {
1330
+ "args": {
1331
+ "should_epoch_stop": false,
1332
+ "should_evaluate": false,
1333
+ "should_log": false,
1334
+ "should_save": true,
1335
+ "should_training_stop": true
1336
+ },
1337
+ "attributes": {}
1338
+ }
1339
+ },
1340
+ "total_flos": 2.5274007305461404e+16,
1341
+ "train_batch_size": 8,
1342
+ "trial_name": null,
1343
+ "trial_params": null
1344
+ }
config.json CHANGED
@@ -7,15 +7,15 @@
7
  "base_model": "jhu-clsp/mmBERT-base",
8
  "bos_token_id": 2,
9
  "category_thresholds": {
10
- "Adversarial": 0.7,
11
- "DirectInjection": 0.4,
12
  "Encoding": 0.45,
13
  "Extraction": 0.55,
14
- "Indirect": 0.2,
15
- "Jailbreak": 0.3,
16
- "Manipulation": 0.4,
17
- "MultiTurn": 0.6,
18
- "Smuggling": 0.6
19
  },
20
  "classifier_activation": "gelu",
21
  "classifier_bias": false,
@@ -46,7 +46,7 @@
46
  "initializer_cutoff_factor": 2.0,
47
  "initializer_range": 0.02,
48
  "intermediate_size": 1152,
49
- "is_valid_threshold": 0.15,
50
  "label2id": {
51
  "Adversarial": 2,
52
  "DirectInjection": 0,
@@ -79,7 +79,7 @@
79
  "sparse_prediction": false,
80
  "training_provenance": {
81
  "base_model": "jhu-clsp/mmBERT-base",
82
- "epochs": 6,
83
  "labels": [
84
  "DirectInjection",
85
  "Jailbreak",
@@ -96,7 +96,7 @@
96
  "problem_type": "multi_label_classification",
97
  "task": "prompt-injection-detection",
98
  "threshold": 0.5,
99
- "trained_at": "2026-06-03T18:16:00+00:00"
100
  },
101
  "transformers_version": "4.57.6",
102
  "vocab_size": 256000
 
7
  "base_model": "jhu-clsp/mmBERT-base",
8
  "bos_token_id": 2,
9
  "category_thresholds": {
10
+ "Adversarial": 0.45,
11
+ "DirectInjection": 0.55,
12
  "Encoding": 0.45,
13
  "Extraction": 0.55,
14
+ "Indirect": 0.25,
15
+ "Jailbreak": 0.05,
16
+ "Manipulation": 0.25,
17
+ "MultiTurn": 0.7,
18
+ "Smuggling": 0.65
19
  },
20
  "classifier_activation": "gelu",
21
  "classifier_bias": false,
 
46
  "initializer_cutoff_factor": 2.0,
47
  "initializer_range": 0.02,
48
  "intermediate_size": 1152,
49
+ "is_valid_threshold": 0.05,
50
  "label2id": {
51
  "Adversarial": 2,
52
  "DirectInjection": 0,
 
79
  "sparse_prediction": false,
80
  "training_provenance": {
81
  "base_model": "jhu-clsp/mmBERT-base",
82
+ "epochs": 10,
83
  "labels": [
84
  "DirectInjection",
85
  "Jailbreak",
 
96
  "problem_type": "multi_label_classification",
97
  "task": "prompt-injection-detection",
98
  "threshold": 0.5,
99
+ "trained_at": "2026-06-03T18:58:34+00:00"
100
  },
101
  "transformers_version": "4.57.6",
102
  "vocab_size": 256000
eval_metrics.json DELETED
@@ -1,40 +0,0 @@
1
- {
2
- "n": 500,
3
- "calibrated": true,
4
- "threshold": "per-class",
5
- "is_valid_threshold": 0.15,
6
- "category_thresholds": {
7
- "DirectInjection": 0.4,
8
- "Jailbreak": 0.3,
9
- "Adversarial": 0.7,
10
- "Extraction": 0.55,
11
- "Encoding": 0.45,
12
- "Manipulation": 0.4,
13
- "Smuggling": 0.6,
14
- "Indirect": 0.2,
15
- "MultiTurn": 0.6
16
- },
17
- "max_seq_length": 3072,
18
- "is_valid_accuracy": 0.948,
19
- "category_set_accuracy": 0.656,
20
- "micro_f1": 0.7768744354110207,
21
- "macro_f1": 0.7687988898685938,
22
- "per_category_f1": {
23
- "DirectInjection": 0.8368794326241135,
24
- "Jailbreak": 0.5964912280701754,
25
- "Adversarial": 0.8368794326241135,
26
- "Extraction": 0.7387387387387387,
27
- "Encoding": 0.8169014084507042,
28
- "Manipulation": 0.631578947368421,
29
- "Smuggling": 0.8793103448275862,
30
- "Indirect": 0.8633093525179856,
31
- "MultiTurn": 0.7191011235955056
32
- },
33
- "latency_ms_per_example": {
34
- "mean": 1.82811776176095,
35
- "p95": 1.953849196434021,
36
- "device": "cuda:0"
37
- },
38
- "base_model": "jhu-clsp/mmBERT-base",
39
- "epochs": 6
40
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
thresholds.json CHANGED
@@ -1,14 +1,14 @@
1
  {
2
  "category_thresholds": {
3
- "DirectInjection": 0.4,
4
- "Jailbreak": 0.3,
5
- "Adversarial": 0.7,
6
  "Extraction": 0.55,
7
  "Encoding": 0.45,
8
- "Manipulation": 0.4,
9
- "Smuggling": 0.6,
10
- "Indirect": 0.2,
11
- "MultiTurn": 0.6
12
  },
13
- "is_valid_threshold": 0.15
14
  }
 
1
  {
2
  "category_thresholds": {
3
+ "DirectInjection": 0.55,
4
+ "Jailbreak": 0.05,
5
+ "Adversarial": 0.45,
6
  "Extraction": 0.55,
7
  "Encoding": 0.45,
8
+ "Manipulation": 0.25,
9
+ "Smuggling": 0.65,
10
+ "Indirect": 0.25,
11
+ "MultiTurn": 0.7
12
  },
13
+ "is_valid_threshold": 0.05
14
  }
training_provenance.json CHANGED
@@ -14,8 +14,8 @@
14
  "MultiTurn"
15
  ],
16
  "max_seq_length": 3072,
17
- "epochs": 6,
18
  "learning_rate": 3e-05,
19
  "threshold": 0.5,
20
- "trained_at": "2026-06-03T18:16:00+00:00"
21
  }
 
14
  "MultiTurn"
15
  ],
16
  "max_seq_length": 3072,
17
+ "epochs": 10,
18
  "learning_rate": 3e-05,
19
  "threshold": 0.5,
20
+ "trained_at": "2026-06-03T18:58:34+00:00"
21
  }