Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/eval_config.yaml +31 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_latest.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_108000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_118647.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_12000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_24000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_36000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_39549.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_48000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_60000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_72000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_79098.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_84000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_96000.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_initial_checkpoint.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_model_best.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_model_final.txt +17 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_latest.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_108000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_118647.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_12000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_24000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_36000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_39549.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_48000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_60000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_72000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_79098.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_84000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_96000.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_initial_checkpoint.txt +3 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_model_best.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_model_final.txt +0 -0
- pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/summary.txt +13 -0
.gitattributes
CHANGED
|
@@ -56,3 +56,4 @@ routing_rerun/N_4.0/wandb/run-20260524_000855-j30yot8x/run-j30yot8x.wandb filter
|
|
| 56 |
routing_rerun/N_6.0/wandb/run-20260524_055222-tjovd8fs/run-tjovd8fs.wandb filter=lfs diff=lfs merge=lfs -text
|
| 57 |
hnet_train_rerun/21-40-57/wandb/run-20260524_214058-3nxcy117/run-3nxcy117.wandb filter=lfs diff=lfs merge=lfs -text
|
| 58 |
hnet_train_rerun/21-40-57/eval_results/predictions_initial_checkpoint.txt filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 56 |
routing_rerun/N_6.0/wandb/run-20260524_055222-tjovd8fs/run-tjovd8fs.wandb filter=lfs diff=lfs merge=lfs -text
|
| 57 |
hnet_train_rerun/21-40-57/wandb/run-20260524_214058-3nxcy117/run-3nxcy117.wandb filter=lfs diff=lfs merge=lfs -text
|
| 58 |
hnet_train_rerun/21-40-57/eval_results/predictions_initial_checkpoint.txt filter=lfs diff=lfs merge=lfs -text
|
| 59 |
+
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_initial_checkpoint.txt filter=lfs diff=lfs merge=lfs -text
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/eval_config.yaml
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
data:
|
| 2 |
+
path: /home/nikita/byte-llms-code//code_completion_exp/datasets/data_V4_full
|
| 3 |
+
max_context_len: 4096
|
| 4 |
+
max_target_len: 256
|
| 5 |
+
num_workers: 4
|
| 6 |
+
pin_memory: true
|
| 7 |
+
max_train_samples: null
|
| 8 |
+
max_val_samples: 2000
|
| 9 |
+
model:
|
| 10 |
+
name: EleutherAI/pythia-1b
|
| 11 |
+
checkpoint_path: null
|
| 12 |
+
from_scratch: false
|
| 13 |
+
paths:
|
| 14 |
+
checkpoints_dir: pythia_1b_rerun/pythia_1b_lr_2e-5
|
| 15 |
+
initial_checkpoint: auto
|
| 16 |
+
output_dir: pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results
|
| 17 |
+
evaluation:
|
| 18 |
+
batch_size: 16
|
| 19 |
+
max_samples: null
|
| 20 |
+
compute_perplexity: true
|
| 21 |
+
bleu_tokenize: none
|
| 22 |
+
save_predictions: true
|
| 23 |
+
use_amp: true
|
| 24 |
+
generation:
|
| 25 |
+
max_new_tokens: 64
|
| 26 |
+
temperature: 0.1
|
| 27 |
+
top_k: 0
|
| 28 |
+
top_p: 1.0
|
| 29 |
+
do_sample: true
|
| 30 |
+
seed: 42
|
| 31 |
+
device: cuda
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_latest.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_latest.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3438372556840846
|
| 5 |
+
token_accuracy: 0.3374259046123755
|
| 6 |
+
bleu: 18.6529732811467
|
| 7 |
+
perplexity: 2.8983173394235715
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 512.0860295599955
|
| 10 |
+
gen_samples_per_s: 68.53926483828818
|
| 11 |
+
gen_time_per_sample_ms: 14.59017692062213
|
| 12 |
+
gen_chars_per_s: 1614.3303122530262
|
| 13 |
+
gen_batch_mean_ms: 661.5939678456297
|
| 14 |
+
gen_batch_p50_ms: 546.2466874998881
|
| 15 |
+
gen_batch_p95_ms: 1306.078345699869
|
| 16 |
+
gen_batch_max_ms: 5513.967688999855
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_108000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_108000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3397629494558094
|
| 5 |
+
token_accuracy: 0.336378091050009
|
| 6 |
+
bleu: 18.52179690907158
|
| 7 |
+
perplexity: 2.9245576470171004
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 528.5562085430115
|
| 10 |
+
gen_samples_per_s: 66.4035336880238
|
| 11 |
+
gen_time_per_sample_ms: 15.059439527694213
|
| 12 |
+
gen_chars_per_s: 1577.5843449053834
|
| 13 |
+
gen_batch_mean_ms: 669.8050106461807
|
| 14 |
+
gen_batch_p50_ms: 554.5250684999701
|
| 15 |
+
gen_batch_p95_ms: 1274.1037359000072
|
| 16 |
+
gen_batch_max_ms: 5511.492358999931
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_118647.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_118647.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.34443558037495015
|
| 5 |
+
token_accuracy: 0.3370600014636126
|
| 6 |
+
bleu: 18.72229382305833
|
| 7 |
+
perplexity: 2.8983173394235715
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 517.0126714489952
|
| 10 |
+
gen_samples_per_s: 67.88615045281828
|
| 11 |
+
gen_time_per_sample_ms: 14.730545086585993
|
| 12 |
+
gen_chars_per_s: 1599.6609864166364
|
| 13 |
+
gen_batch_mean_ms: 663.5231613442652
|
| 14 |
+
gen_batch_p50_ms: 551.6623404996608
|
| 15 |
+
gen_batch_p95_ms: 1254.7493076996939
|
| 16 |
+
gen_batch_max_ms: 5509.677142999863
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_12000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_12000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.31092939768647787
|
| 5 |
+
token_accuracy: 0.3056821432610619
|
| 6 |
+
bleu: 16.91287733900013
|
| 7 |
+
perplexity: 3.0910473571744728
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 489.6108111139956
|
| 10 |
+
gen_samples_per_s: 71.68550857801252
|
| 11 |
+
gen_time_per_sample_ms: 13.949820819248835
|
| 12 |
+
gen_chars_per_s: 1568.268066329982
|
| 13 |
+
gen_batch_mean_ms: 640.881225442608
|
| 14 |
+
gen_batch_p50_ms: 530.8562299996993
|
| 15 |
+
gen_batch_p95_ms: 1323.3005196497725
|
| 16 |
+
gen_batch_max_ms: 3567.599123999571
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_24000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_24000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3176819191976751
|
| 5 |
+
token_accuracy: 0.31616360527698867
|
| 6 |
+
bleu: 16.999023276021926
|
| 7 |
+
perplexity: 3.049432811892437
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 517.3736476249842
|
| 10 |
+
gen_samples_per_s: 67.838785684424
|
| 11 |
+
gen_time_per_sample_ms: 14.740829894153062
|
| 12 |
+
gen_chars_per_s: 1535.665768148862
|
| 13 |
+
gen_batch_mean_ms: 669.2715530560091
|
| 14 |
+
gen_batch_p50_ms: 551.570622500094
|
| 15 |
+
gen_batch_p95_ms: 1301.835535300187
|
| 16 |
+
gen_batch_max_ms: 5511.411889999181
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_36000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_36000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3228389081998974
|
| 5 |
+
token_accuracy: 0.31837898252313507
|
| 6 |
+
bleu: 17.449522226725577
|
| 7 |
+
perplexity: 2.9921718581044807
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 517.3814206280158
|
| 10 |
+
gen_samples_per_s: 67.83776649226563
|
| 11 |
+
gen_time_per_sample_ms: 14.741051359850013
|
| 12 |
+
gen_chars_per_s: 1562.0835379418666
|
| 13 |
+
gen_batch_mean_ms: 686.4863323565612
|
| 14 |
+
gen_batch_p50_ms: 547.4083809999684
|
| 15 |
+
gen_batch_p95_ms: 1377.9240408999108
|
| 16 |
+
gen_batch_max_ms: 5507.957126999827
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_39549.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_39549.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.32688472277622654
|
| 5 |
+
token_accuracy: 0.32353489052843065
|
| 6 |
+
bleu: 17.599796716522864
|
| 7 |
+
perplexity: 2.9765649173167703
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 535.7109377580045
|
| 10 |
+
gen_samples_per_s: 65.51667611433899
|
| 11 |
+
gen_time_per_sample_ms: 15.26328958225553
|
| 12 |
+
gen_chars_per_s: 1543.2072443020552
|
| 13 |
+
gen_batch_mean_ms: 681.6499375327907
|
| 14 |
+
gen_batch_p50_ms: 560.1317364998977
|
| 15 |
+
gen_batch_p95_ms: 1265.995824399715
|
| 16 |
+
gen_batch_max_ms: 5511.29786599995
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_48000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_48000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3267992478203886
|
| 5 |
+
token_accuracy: 0.32422345372655725
|
| 6 |
+
bleu: 16.98040900700143
|
| 7 |
+
perplexity: 3.000834224455479
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 552.924158363734
|
| 10 |
+
gen_samples_per_s: 63.47706004357877
|
| 11 |
+
gen_time_per_sample_ms: 15.75372267262334
|
| 12 |
+
gen_chars_per_s: 1491.5879285156116
|
| 13 |
+
gen_batch_mean_ms: 498.38210237196296
|
| 14 |
+
gen_batch_p50_ms: 425.19649300083984
|
| 15 |
+
gen_batch_p95_ms: 1108.712311401905
|
| 16 |
+
gen_batch_max_ms: 3580.2823460035142
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_60000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_60000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.33198472847455696
|
| 5 |
+
token_accuracy: 0.31948999753846974
|
| 6 |
+
bleu: 18.78148144552106
|
| 7 |
+
perplexity: 2.993654101091439
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 521.2763403879362
|
| 10 |
+
gen_samples_per_s: 67.33089012610837
|
| 11 |
+
gen_time_per_sample_ms: 14.852024058007187
|
| 12 |
+
gen_chars_per_s: 1513.8035987068602
|
| 13 |
+
gen_batch_mean_ms: 461.5683323910272
|
| 14 |
+
gen_batch_p50_ms: 406.8481369977235
|
| 15 |
+
gen_batch_p95_ms: 841.7265435971791
|
| 16 |
+
gen_batch_max_ms: 3245.230978995096
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_72000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_72000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3257165650464414
|
| 5 |
+
token_accuracy: 0.32552407310079634
|
| 6 |
+
bleu: 17.554073235328627
|
| 7 |
+
perplexity: 2.9570823713233994
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 556.0231101891404
|
| 10 |
+
gen_samples_per_s: 63.123275556048085
|
| 11 |
+
gen_time_per_sample_ms: 15.842016929430178
|
| 12 |
+
gen_chars_per_s: 1494.0727908187314
|
| 13 |
+
gen_batch_mean_ms: 486.72739216855683
|
| 14 |
+
gen_batch_p50_ms: 430.15693000052124
|
| 15 |
+
gen_batch_p95_ms: 946.6991030014475
|
| 16 |
+
gen_batch_max_ms: 3250.2466439982527
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_79098.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_79098.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.32924952988774286
|
| 5 |
+
token_accuracy: 0.32587001789599035
|
| 6 |
+
bleu: 17.66041023077036
|
| 7 |
+
perplexity: 2.9432653848024013
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 558.4788690531059
|
| 10 |
+
gen_samples_per_s: 62.84570812770093
|
| 11 |
+
gen_time_per_sample_ms: 15.91198555624554
|
| 12 |
+
gen_chars_per_s: 1479.9951185286538
|
| 13 |
+
gen_batch_mean_ms: 481.4050168861154
|
| 14 |
+
gen_batch_p50_ms: 412.99323799466947
|
| 15 |
+
gen_batch_p95_ms: 985.8789092031672
|
| 16 |
+
gen_batch_max_ms: 3919.5375990020693
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_84000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_84000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.3267992478203886
|
| 5 |
+
token_accuracy: 0.32729704017616573
|
| 6 |
+
bleu: 17.15454805989793
|
| 7 |
+
perplexity: 3.039049060843566
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 554.8222870400496
|
| 10 |
+
gen_samples_per_s: 63.25989568163557
|
| 11 |
+
gen_time_per_sample_ms: 15.807803494217607
|
| 12 |
+
gen_chars_per_s: 1502.05573832661
|
| 13 |
+
gen_batch_mean_ms: 499.97872161993917
|
| 14 |
+
gen_batch_p50_ms: 430.8593480018317
|
| 15 |
+
gen_batch_p95_ms: 1011.4186708029592
|
| 16 |
+
gen_batch_max_ms: 3927.7867519995198
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_checkpoint_step_96000.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: checkpoint_step_96000.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.32765399737876805
|
| 5 |
+
token_accuracy: 0.32871408327955665
|
| 6 |
+
bleu: 17.456606748558585
|
| 7 |
+
perplexity: 3.008593212810714
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 569.5983627550886
|
| 10 |
+
gen_samples_per_s: 61.61885689108127
|
| 11 |
+
gen_time_per_sample_ms: 16.228798300617942
|
| 12 |
+
gen_chars_per_s: 1483.8876922183517
|
| 13 |
+
gen_batch_mean_ms: 500.05995249138255
|
| 14 |
+
gen_batch_p50_ms: 431.5248690036242
|
| 15 |
+
gen_batch_p95_ms: 1040.302878398508
|
| 16 |
+
gen_batch_max_ms: 3247.8538199939067
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_initial_checkpoint.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: initial_checkpoint
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.0
|
| 5 |
+
token_accuracy: 0.270582052117914
|
| 6 |
+
bleu: 0.9973597289127779
|
| 7 |
+
perplexity: 11.896337297460965
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 1114.1351297710041
|
| 10 |
+
gen_samples_per_s: 31.502462369366214
|
| 11 |
+
gen_time_per_sample_ms: 31.74355033822452
|
| 12 |
+
gen_chars_per_s: 6894.466204991493
|
| 13 |
+
gen_batch_mean_ms: 1458.6486288237702
|
| 14 |
+
gen_batch_p50_ms: 1228.9804150000236
|
| 15 |
+
gen_batch_p95_ms: 2486.2865908000094
|
| 16 |
+
gen_batch_max_ms: 5505.641633999858
|
| 17 |
+
gen_num_batches: 732
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_model_best.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: model_best.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.33959199954413355
|
| 5 |
+
token_accuracy: 0.33668079274580376
|
| 6 |
+
bleu: 18.43932275787846
|
| 7 |
+
perplexity: 2.9246050551995744
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 551.2998683829064
|
| 10 |
+
gen_samples_per_s: 63.664081950447
|
| 11 |
+
gen_time_per_sample_ms: 15.707443967830258
|
| 12 |
+
gen_chars_per_s: 1512.6196246809336
|
| 13 |
+
gen_batch_mean_ms: 486.2616247575008
|
| 14 |
+
gen_batch_p50_ms: 426.6929829973378
|
| 15 |
+
gen_batch_p95_ms: 928.941303402826
|
| 16 |
+
gen_batch_max_ms: 3925.172384995676
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/metrics_model_final.txt
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Checkpoint: model_final.pt
|
| 2 |
+
================================================================================
|
| 3 |
+
|
| 4 |
+
exact_match: 0.34412217220354435
|
| 5 |
+
token_accuracy: 0.33789825231350584
|
| 6 |
+
bleu: 18.793073765848895
|
| 7 |
+
perplexity: 2.898342720850918
|
| 8 |
+
num_samples: 35098
|
| 9 |
+
gen_wall_time_s: 549.0805001499248
|
| 10 |
+
gen_samples_per_s: 63.92141041325743
|
| 11 |
+
gen_time_per_sample_ms: 15.64421050059618
|
| 12 |
+
gen_chars_per_s: 1505.380358206686
|
| 13 |
+
gen_batch_mean_ms: 483.941579053738
|
| 14 |
+
gen_batch_p50_ms: 425.6622940010857
|
| 15 |
+
gen_batch_p95_ms: 903.3644577983071
|
| 16 |
+
gen_batch_max_ms: 4074.6655790062505
|
| 17 |
+
gen_num_batches: 1097
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_latest.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_108000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_118647.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_12000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_24000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_36000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_39549.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_48000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_60000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_72000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_79098.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_84000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_checkpoint_step_96000.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_initial_checkpoint.txt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1e88aa8613bf304363046c6536a5d25a53fa94435859068e58c383c26e080abe
|
| 3 |
+
size 15725651
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_model_best.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/predictions_model_final.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pythia_1b_rerun/pythia_1b_lr_2e-5/eval_results_full_updated/summary.txt
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
EVALUATION SUMMARY
|
| 2 |
+
==================================================================================================
|
| 3 |
+
|
| 4 |
+
Checkpoint Exact Match Token Acc BLEU PERPLEXITY ms/sample samp/s
|
| 5 |
+
--------------------------------------------------------------------------------------------------
|
| 6 |
+
checkpoint_step_48000 32.68% 32.42% 16.98 3.00 15.8 63.48
|
| 7 |
+
checkpoint_step_60000 33.20% 31.95% 18.78 2.99 14.9 67.33
|
| 8 |
+
checkpoint_step_72000 32.57% 32.55% 17.55 2.96 15.8 63.12
|
| 9 |
+
checkpoint_step_79098 32.92% 32.59% 17.66 2.94 15.9 62.85
|
| 10 |
+
checkpoint_step_84000 32.68% 32.73% 17.15 3.04 15.8 63.26
|
| 11 |
+
checkpoint_step_96000 32.77% 32.87% 17.46 3.01 16.2 61.62
|
| 12 |
+
model_best 33.96% 33.67% 18.44 2.92 15.7 63.66
|
| 13 |
+
model_final 34.41% 33.79% 18.79 2.90 15.6 63.92
|