datasysdev commited on
Commit
faa1ec5
·
verified ·
1 Parent(s): f43f277

Update logs/k_sweep_all32_step1000_exact.log

Browse files
logs/k_sweep_all32_step1000_exact.log ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2
+ Loading base model Qwen/Qwen3-4B-Instruct-2507 ...
3
+
4
+ Loaded ckpt step 1000
5
+ Pre-running teacher captures...
6
+ Computing full-attention PPL...
7
+ ppl_full = 20.5349
8
+
9
+ === K = 16 ===
10
+ mass_avg = 0.5457 recall_avg = 0.5178 ppl_ann = 24.8603 ppl_gap = +21.064%
11
+
12
+ === K = 32 ===
13
+ mass_avg = 0.6270 recall_avg = 0.5719 ppl_ann = 21.8537 ppl_gap = +6.422%
14
+
15
+ === K = 64 ===
16
+ mass_avg = 0.7224 recall_avg = 0.6515 ppl_ann = 20.9403 ppl_gap = +1.974%
17
+
18
+ === K = 128 ===
19
+ mass_avg = 0.8071 recall_avg = 0.7456 ppl_ann = 20.6561 ppl_gap = +0.590%
20
+
21
+ === K = 256 ===
22
+ mass_avg = 0.9024 recall_avg = 0.8756 ppl_ann = 20.5223 ppl_gap = -0.062%
23
+
24
+ === K = 512 ===
25
+ mass_avg = 0.0000 recall_avg = 0.0000 ppl_ann = 20.5214 ppl_gap = -0.066%
26
+
27
+ Wrote /tmp/checkpoints_all32_d128_block_reserve_0_1_2_35/search_step_1000.k_sweep_exact.json