LogicPalette commited on
Commit
f741e18
·
verified ·
1 Parent(s): 70c5d01

data update

Browse files
audit/dbs.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:bf34c4840018c88f92929a0b453588edd28b8eb2217cfe9700d8519ada63ce81
3
  size 23599633
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:048d734b24079c0d9e396af2d2cf8f26c938a8f6852e7e7e84f5fc6ad62dd7c9
3
  size 23599633
results/e4_eval.json CHANGED
@@ -1,4 +1,122 @@
1
  {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  "base": {
3
  "items": 331,
4
  "samples": 2648,
@@ -58,6 +176,63 @@
58
  "4": 4.3159257660768235e-05
59
  }
60
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
61
  "sft": {
62
  "items": 331,
63
  "samples": 2648,
@@ -176,66 +351,22 @@
176
  "4": 0.021450151057401813
177
  }
178
  },
179
- "B": {
180
- "items": 331,
181
- "samples": 2648,
182
- "verified": [
183
- 0.1268882175226586,
184
- 0.11474862442249444,
185
- 0.1401088301102348
186
- ],
187
- "single": [
188
- 0.1782477341389728,
189
- 0.16413956667413937,
190
- 0.19328811678930724
191
- ],
192
- "silent_error": [
193
- 0.6404833836858006,
194
- 0.6220147464941264,
195
- 0.6585449973457771
196
- ],
197
- "smuggled": [
198
- 0.0,
199
- 0.0,
200
- 0.0014486536450744267
201
- ],
202
- "has_sql": [
203
- 0.9879154078549849,
204
- 0.9829902434304691,
205
- 0.9914269314113464
206
- ],
207
- "inflation_single_minus_verified": 0.0513595166163142,
208
- "invalid_tool_call_rate": 0.16150956895009838,
209
- "tool_calls_per_traj": 2.111404833836858,
210
- "verified_by_family": {
211
- "F1": [
212
- 0.5659722222222222,
213
- 0.5082329974060685,
214
- 0.6219746220697019
215
- ],
216
- "F2": [
217
- 0.010416666666666666,
218
- 0.0035487456405297844,
219
- 0.03017365719717396
220
- ],
221
- "F3": [
222
- 0.0,
223
- 0.0,
224
- 0.010558440816003448
225
  ],
226
- "F4": [
227
- 0.09929906542056074,
228
- 0.08601700626020076,
229
- 0.11437538427281815
 
230
  ]
231
  },
232
- "pass^k": {
233
- "1": 0.1268882175226586,
234
- "2": 0.0807078118256366,
235
- "4": 0.03927492447129909
236
- }
237
- },
238
- "comparisons": {
239
  "B_vs_sft_ho": {
240
  "n_items": 331,
241
  "mcnemar_p_sample0": 0.006426576095691416,
@@ -243,6 +374,27 @@
243
  0.0324773413897281,
244
  0.01812688821752266,
245
  0.047960725075528704
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
246
  ]
247
  },
248
  "sft_ho_vs_base": {
@@ -252,6 +404,12 @@
252
  0.08496978851963746,
253
  0.06306646525679759,
254
  0.10725075528700906
 
 
 
 
 
 
255
  ]
256
  },
257
  "sft_vs_base": {
@@ -261,6 +419,12 @@
261
  0.10800604229607251,
262
  0.08270392749244714,
263
  0.13481873111782477
 
 
 
 
 
 
264
  ]
265
  },
266
  "sft_vs_sft_ho": {
@@ -270,6 +434,57 @@
270
  0.023036253776435044,
271
  0.0007552870090634441,
272
  0.04607250755287009
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
273
  ]
274
  }
275
  }
 
1
  {
2
+ "B": {
3
+ "items": 331,
4
+ "samples": 2648,
5
+ "verified": [
6
+ 0.1268882175226586,
7
+ 0.11474862442249444,
8
+ 0.1401088301102348
9
+ ],
10
+ "single": [
11
+ 0.1782477341389728,
12
+ 0.16413956667413937,
13
+ 0.19328811678930724
14
+ ],
15
+ "silent_error": [
16
+ 0.6404833836858006,
17
+ 0.6220147464941264,
18
+ 0.6585449973457771
19
+ ],
20
+ "smuggled": [
21
+ 0.0,
22
+ 0.0,
23
+ 0.0014486536450744267
24
+ ],
25
+ "has_sql": [
26
+ 0.9879154078549849,
27
+ 0.9829902434304691,
28
+ 0.9914269314113464
29
+ ],
30
+ "inflation_single_minus_verified": 0.0513595166163142,
31
+ "invalid_tool_call_rate": 0.16150956895009838,
32
+ "tool_calls_per_traj": 2.111404833836858,
33
+ "verified_by_family": {
34
+ "F1": [
35
+ 0.5659722222222222,
36
+ 0.5082329974060685,
37
+ 0.6219746220697019
38
+ ],
39
+ "F2": [
40
+ 0.010416666666666666,
41
+ 0.0035487456405297844,
42
+ 0.03017365719717396
43
+ ],
44
+ "F3": [
45
+ 0.0,
46
+ 0.0,
47
+ 0.010558440816003448
48
+ ],
49
+ "F4": [
50
+ 0.09929906542056074,
51
+ 0.08601700626020076,
52
+ 0.11437538427281815
53
+ ]
54
+ },
55
+ "pass^k": {
56
+ "1": 0.1268882175226586,
57
+ "2": 0.0807078118256366,
58
+ "4": 0.03927492447129909
59
+ }
60
+ },
61
+ "C": {
62
+ "items": 331,
63
+ "samples": 2648,
64
+ "verified": [
65
+ 0.08950151057401813,
66
+ 0.07921473537422782,
67
+ 0.10097762603981741
68
+ ],
69
+ "single": [
70
+ 0.1404833836858006,
71
+ 0.12776814975928472,
72
+ 0.15424024772569325
73
+ ],
74
+ "silent_error": [
75
+ 0.6117824773413897,
76
+ 0.5930709276238418,
77
+ 0.6301701588724254
78
+ ],
79
+ "smuggled": [
80
+ 0.0,
81
+ 0.0,
82
+ 0.0014486536450744267
83
+ ],
84
+ "has_sql": [
85
+ 0.9898036253776435,
86
+ 0.9852051270532216,
87
+ 0.9929830120875174
88
+ ],
89
+ "inflation_single_minus_verified": 0.050981873111782476,
90
+ "invalid_tool_call_rate": 0.19475857117214146,
91
+ "tool_calls_per_traj": 2.1038519637462234,
92
+ "verified_by_family": {
93
+ "F1": [
94
+ 0.2881944444444444,
95
+ 0.23894335969829233,
96
+ 0.3430216514584447
97
+ ],
98
+ "F2": [
99
+ 0.013888888888888888,
100
+ 0.005413932945500995,
101
+ 0.03516150249619066
102
+ ],
103
+ "F3": [
104
+ 0.0,
105
+ 0.0,
106
+ 0.010558440816003448
107
+ ],
108
+ "F4": [
109
+ 0.08761682242990654,
110
+ 0.07513001880538675,
111
+ 0.10195019652451318
112
+ ]
113
+ },
114
+ "pass^k": {
115
+ "1": 0.08950151057401813,
116
+ "2": 0.04650410012947777,
117
+ "4": 0.021881743634009496
118
+ }
119
+ },
120
  "base": {
121
  "items": 331,
122
  "samples": 2648,
 
176
  "4": 4.3159257660768235e-05
177
  }
178
  },
179
+ "command_a": {
180
+ "items": 331,
181
+ "samples": 331,
182
+ "verified": [
183
+ 0.8096676737160121,
184
+ 0.7639169970960494,
185
+ 0.8483127855832832
186
+ ],
187
+ "single": [
188
+ 0.8277945619335347,
189
+ 0.7834184000616147,
190
+ 0.8646492235550447
191
+ ],
192
+ "silent_error": [
193
+ 0.0,
194
+ 0.0,
195
+ 0.011472887478736215
196
+ ],
197
+ "smuggled": [
198
+ 0.0,
199
+ 0.0,
200
+ 0.011472887478736215
201
+ ],
202
+ "has_sql": [
203
+ 1.0,
204
+ 0.9885271125212638,
205
+ 1.0
206
+ ],
207
+ "inflation_single_minus_verified": 0.01812688821752266,
208
+ "invalid_tool_call_rate": null,
209
+ "tool_calls_per_traj": 0.0,
210
+ "verified_by_family": {
211
+ "F1": [
212
+ 1.0,
213
+ 0.9035781695514236,
214
+ 1.0
215
+ ],
216
+ "F2": [
217
+ 0.0,
218
+ 0.0,
219
+ 0.09642183044857634
220
+ ],
221
+ "F3": [
222
+ 0.6444444444444445,
223
+ 0.49835546011952203,
224
+ 0.7678110864309596
225
+ ],
226
+ "F4": [
227
+ 0.9485981308411215,
228
+ 0.9103152736698502,
229
+ 0.97105908274325
230
+ ]
231
+ },
232
+ "pass^k": {
233
+ "1": 0.8096676737160121
234
+ }
235
+ },
236
  "sft": {
237
  "items": 331,
238
  "samples": 2648,
 
351
  "4": 0.021450151057401813
352
  }
353
  },
354
+ "comparisons": {
355
+ "C_vs_B": {
356
+ "n_items": 331,
357
+ "mcnemar_p_sample0": 7.025478407740593e-05,
358
+ "verified_diff_bootstrap95": [
359
+ -0.037386706948640484,
360
+ -0.05324773413897281,
361
+ -0.022658610271903322
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
362
  ],
363
+ "silent_mcnemar_p_sample0": 0.606295937693966,
364
+ "silent_error_diff_bootstrap95": [
365
+ -0.028700906344410877,
366
+ -0.052870090634441085,
367
+ -0.005287009063444109
368
  ]
369
  },
 
 
 
 
 
 
 
370
  "B_vs_sft_ho": {
371
  "n_items": 331,
372
  "mcnemar_p_sample0": 0.006426576095691416,
 
374
  0.0324773413897281,
375
  0.01812688821752266,
376
  0.047960725075528704
377
+ ],
378
+ "silent_mcnemar_p_sample0": 0.06138589870445953,
379
+ "silent_error_diff_bootstrap95": [
380
+ -0.013972809667673716,
381
+ -0.03625377643504532,
382
+ 0.008685800604229608
383
+ ]
384
+ },
385
+ "C_vs_sft_ho": {
386
+ "n_items": 331,
387
+ "mcnemar_p_sample0": 0.21003961563110352,
388
+ "verified_diff_bootstrap95": [
389
+ -0.004909365558912387,
390
+ -0.015483383685800604,
391
+ 0.0056646525679758305
392
+ ],
393
+ "silent_mcnemar_p_sample0": 0.1979791406725175,
394
+ "silent_error_diff_bootstrap95": [
395
+ -0.04267371601208459,
396
+ -0.06570996978851963,
397
+ -0.02039274924471299
398
  ]
399
  },
400
  "sft_ho_vs_base": {
 
404
  0.08496978851963746,
405
  0.06306646525679759,
406
  0.10725075528700906
407
+ ],
408
+ "silent_mcnemar_p_sample0": 5.104817220120468e-61,
409
+ "silent_error_diff_bootstrap95": [
410
+ 0.6095166163141994,
411
+ 0.574773413897281,
412
+ 0.6450151057401813
413
  ]
414
  },
415
  "sft_vs_base": {
 
419
  0.10800604229607251,
420
  0.08270392749244714,
421
  0.13481873111782477
422
+ ],
423
+ "silent_mcnemar_p_sample0": 3.838328256567196e-49,
424
+ "silent_error_diff_bootstrap95": [
425
+ 0.5868580060422961,
426
+ 0.5513595166163142,
427
+ 0.622356495468278
428
  ]
429
  },
430
  "sft_vs_sft_ho": {
 
434
  0.023036253776435044,
435
  0.0007552870090634441,
436
  0.04607250755287009
437
+ ],
438
+ "silent_mcnemar_p_sample0": 0.2066471089849431,
439
+ "silent_error_diff_bootstrap95": [
440
+ -0.022658610271903322,
441
+ -0.052870090634441085,
442
+ 0.008308157099697885
443
+ ]
444
+ },
445
+ "C_vs_command_a": {
446
+ "n_items": 331,
447
+ "mcnemar_p_sample0": 8.843436600416711e-75,
448
+ "verified_diff_bootstrap95": [
449
+ -0.720166163141994,
450
+ -0.7628398791540786,
451
+ -0.6763595166163142
452
+ ],
453
+ "silent_mcnemar_p_sample0": 4.861730685829017e-63,
454
+ "silent_error_diff_bootstrap95": [
455
+ 0.6117824773413897,
456
+ 0.573262839879154,
457
+ 0.6487915407854985
458
+ ]
459
+ },
460
+ "B_vs_command_a": {
461
+ "n_items": 331,
462
+ "mcnemar_p_sample0": 2.967364920549937e-67,
463
+ "verified_diff_bootstrap95": [
464
+ -0.6827794561933535,
465
+ -0.7265861027190332,
466
+ -0.6374622356495468
467
+ ],
468
+ "silent_mcnemar_p_sample0": 3.111507638930571e-61,
469
+ "silent_error_diff_bootstrap95": [
470
+ 0.6404833836858006,
471
+ 0.6049848942598187,
472
+ 0.675226586102719
473
+ ]
474
+ },
475
+ "sft_ho_vs_command_a": {
476
+ "n_items": 331,
477
+ "mcnemar_p_sample0": 6.876656300484035e-71,
478
+ "verified_diff_bootstrap95": [
479
+ -0.7152567975830816,
480
+ -0.7579305135951662,
481
+ -0.6714501510574018
482
+ ],
483
+ "silent_mcnemar_p_sample0": 5.934729841099874e-67,
484
+ "silent_error_diff_bootstrap95": [
485
+ 0.6544561933534743,
486
+ 0.6204682779456193,
487
+ 0.6884441087613293
488
  ]
489
  }
490
  }
results/e5_roi.json ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "command_a": {
3
+ "n": 331,
4
+ "verified": [
5
+ 0.8096676737160121,
6
+ 0.7639169970960494,
7
+ 0.8483127855832832
8
+ ],
9
+ "single": [
10
+ 0.8277945619335347,
11
+ 0.7834184000616147,
12
+ 0.8646492235550447
13
+ ],
14
+ "usd_per_1k": 4.244501510574018,
15
+ "tokens_per_query": [
16
+ 666.404833836858,
17
+ 257.8489425981873
18
+ ],
19
+ "latency_s_median": 9.74,
20
+ "latency_s_p90": 42.22
21
+ },
22
+ "sft_ho": {
23
+ "n": 2648,
24
+ "verified": [
25
+ 0.09441087613293052,
26
+ 0.08385386452215371,
27
+ 0.10614300406909247
28
+ ],
29
+ "single": [
30
+ 0.145392749244713,
31
+ 0.13248019213082912,
32
+ 0.15933271253134978
33
+ ],
34
+ "gen_s": 465.57,
35
+ "trajectories_per_s": 5.687651695770776,
36
+ "usd_per_1k": 0.03614067220543806
37
+ },
38
+ "B": {
39
+ "n": 2648,
40
+ "verified": [
41
+ 0.1268882175226586,
42
+ 0.11474862442249444,
43
+ 0.1401088301102348
44
+ ],
45
+ "single": [
46
+ 0.1782477341389728,
47
+ 0.16413956667413937,
48
+ 0.19328811678930724
49
+ ],
50
+ "gen_s": 523.65,
51
+ "trajectories_per_s": 5.056812756612241,
52
+ "usd_per_1k": 0.04064923212487412
53
+ },
54
+ "C": {
55
+ "n": 2648,
56
+ "verified": [
57
+ 0.08950151057401813,
58
+ 0.07921473537422782,
59
+ 0.10097762603981741
60
+ ],
61
+ "single": [
62
+ 0.1404833836858006,
63
+ 0.12776814975928472,
64
+ 0.15424024772569325
65
+ ],
66
+ "gen_s": 443.21,
67
+ "trajectories_per_s": 5.974594436046118,
68
+ "usd_per_1k": 0.034404938737831486
69
+ },
70
+ "assumptions": {
71
+ "command_a_usd_per_1M": [
72
+ 2.5,
73
+ 10.0
74
+ ],
75
+ "gpu_usd_hr": 0.74,
76
+ "throughput_gpu": "RTX 5070 Ti laptop, vLLM batched; latency per query not measured for Tiny Aya"
77
+ }
78
+ }
results/figure3_training_rewards.json CHANGED
@@ -5,5 +5,12 @@
5
  "lucky_share_of_passes": 0.0819935691318328,
6
  "lucky_share_first_half": 0.10666666666666667,
7
  "lucky_share_second_half": 0.059006211180124224
 
 
 
 
 
 
 
8
  }
9
  }
 
5
  "lucky_share_of_passes": 0.0819935691318328,
6
  "lucky_share_first_half": 0.10666666666666667,
7
  "lucky_share_second_half": 0.059006211180124224
8
+ },
9
+ "C": {
10
+ "steps": 60,
11
+ "single_pass": 0.8652777777777778,
12
+ "lucky_share_of_passes": 0.08346709470304976,
13
+ "lucky_share_first_half": 0.11147540983606558,
14
+ "lucky_share_second_half": 0.05660377358490566
15
  }
16
  }
results/figure3_training_rewards.png CHANGED

Git LFS Details

  • SHA256: 308d3a551987cdfc19cb5c8edd5bd70764f05e80f3e0364c331f8363aa74e40e
  • Pointer size: 131 Bytes
  • Size of remote file: 116 kB

Git LFS Details

  • SHA256: 602c1e010ccfd9236c31ff8099172b2aeb760349b7945b533307780c82201d24
  • Pointer size: 131 Bytes
  • Size of remote file: 144 kB
results/traj_C.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
results/traj_command_a.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
runs/grpo_C/grpo_training_config.yaml ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ accelerator_config:
2
+ dispatch_batches: null
3
+ even_batches: true
4
+ gradient_accumulation_kwargs: null
5
+ non_blocking: false
6
+ split_batches: false
7
+ use_seedable_sampler: true
8
+ adam_beta1: 0.9
9
+ adam_beta2: 0.999
10
+ adam_epsilon: 1.0e-08
11
+ auto_find_batch_size: false
12
+ average_tokens_across_devices: true
13
+ batch_eval_metrics: false
14
+ beta: 0.02
15
+ bf16: true
16
+ bf16_full_eval: false
17
+ cache_implementation: null
18
+ cast_lm_head_to_fp32: false
19
+ chat_template_kwargs: null
20
+ data_seed: null
21
+ dataloader_drop_last: false
22
+ dataloader_in_order: true
23
+ dataloader_multiprocessing_context: null
24
+ dataloader_num_workers: 0
25
+ dataloader_persistent_workers: false
26
+ dataloader_pin_memory: true
27
+ dataloader_prefetch_factor: null
28
+ ddp_backend: null
29
+ ddp_broadcast_buffers: null
30
+ ddp_bucket_cap_mb: null
31
+ ddp_find_unused_parameters: null
32
+ ddp_static_graph: null
33
+ ddp_timeout: 1800
34
+ debug: []
35
+ deepspeed: null
36
+ delta: null
37
+ disable_dropout: false
38
+ disable_tqdm: false
39
+ do_eval: false
40
+ do_predict: false
41
+ do_train: false
42
+ ds3_gather_for_generation: true
43
+ enable_jit_checkpoint: false
44
+ epsilon: 0.2
45
+ epsilon_high: null
46
+ eval_accumulation_steps: null
47
+ eval_delay: 0
48
+ eval_do_concat_batches: true
49
+ eval_on_start: false
50
+ eval_steps: null
51
+ eval_strategy: 'no'
52
+ eval_use_gather_object: false
53
+ fp16: false
54
+ fp16_full_eval: false
55
+ fsdp: null
56
+ fsdp_config: null
57
+ full_determinism: false
58
+ generation_batch_size: 24
59
+ generation_kwargs: null
60
+ gradient_accumulation_steps: 24
61
+ gradient_checkpointing: true
62
+ gradient_checkpointing_kwargs: null
63
+ greater_is_better: null
64
+ hub_always_push: false
65
+ hub_model_id: null
66
+ hub_private_repo: null
67
+ hub_revision: null
68
+ hub_strategy: every_save
69
+ hub_token: <HUB_TOKEN>
70
+ ignore_data_skip: false
71
+ importance_sampling_level: token
72
+ include_for_metrics: []
73
+ include_num_input_tokens_seen: 'no'
74
+ label_names: null
75
+ label_smoothing_factor: 0.0
76
+ learning_rate: 1.0e-05
77
+ length_column_name: length
78
+ liger_kernel_config: null
79
+ load_best_model_at_end: false
80
+ local_rank: -1
81
+ log_completions: false
82
+ log_completions_hub_repo: null
83
+ log_level: passive
84
+ log_level_replica: warning
85
+ log_on_each_node: true
86
+ log_unique_prompts: false
87
+ logging_first_step: false
88
+ logging_nan_inf_filter: true
89
+ logging_steps: 1
90
+ logging_strategy: steps
91
+ loss_type: dapo
92
+ lr_scheduler_kwargs: null
93
+ lr_scheduler_type: linear
94
+ mask_truncated_completions: false
95
+ max_completion_length: 2048
96
+ max_grad_norm: 1.0
97
+ max_steps: 60
98
+ max_tool_calling_iterations: null
99
+ metric_for_best_model: null
100
+ min_p: null
101
+ model_init_kwargs: null
102
+ multi_objective_aggregation: sum_then_normalize
103
+ neftune_noise_alpha: null
104
+ num_completions_to_print: null
105
+ num_generations: 6
106
+ num_generations_eval: null
107
+ num_iterations: 1
108
+ num_train_epochs: 1
109
+ off_policy_mask_threshold: null
110
+ optim: adamw_torch_fused
111
+ optim_args: null
112
+ optim_target_modules: null
113
+ output_dir: runs/grpo_C
114
+ pad_to_multiple_of: null
115
+ parallelism_config: null
116
+ per_device_eval_batch_size: 8
117
+ per_device_train_batch_size: 1
118
+ prediction_loss_only: false
119
+ project: huggingface
120
+ push_to_hub: false
121
+ ref_model_mixup_alpha: 0.6
122
+ ref_model_sync_steps: 512
123
+ remove_unused_columns: false
124
+ repetition_penalty: 1.0
125
+ report_to: []
126
+ restore_callback_states_from_checkpoint: false
127
+ resume_from_checkpoint: null
128
+ reward_weights: null
129
+ router_aux_loss_coef: 0.001
130
+ run_name: null
131
+ sapo_temperature_neg: 1.05
132
+ sapo_temperature_pos: 1.0
133
+ save_on_each_node: false
134
+ save_only_model: false
135
+ save_steps: 20
136
+ save_strategy: steps
137
+ save_total_limit: null
138
+ scale_rewards: group
139
+ seed: 0
140
+ shuffle_dataset: true
141
+ skip_memory_metrics: true
142
+ steps_per_generation: 24
143
+ sync_ref_model: false
144
+ temperature: 1.0
145
+ tf32: null
146
+ top_entropy_quantile: 1.0
147
+ top_k: 0
148
+ top_p: 0.95
149
+ torch_compile: false
150
+ torch_compile_backend: null
151
+ torch_compile_mode: null
152
+ torch_empty_cache_steps: null
153
+ trackio_bucket_id: null
154
+ trackio_space_id: null
155
+ trackio_static_space_id: null
156
+ train_sampling_strategy: random
157
+ transformers_continuous_batching_config: null
158
+ trust_remote_code: false
159
+ use_bias_correction_kl: false
160
+ use_cache: false
161
+ use_cpu: false
162
+ use_liger_kernel: false
163
+ use_transformers_continuous_batching: false
164
+ use_transformers_paged: false
165
+ use_vllm: true
166
+ vespo_k_neg: 3.0
167
+ vespo_k_pos: 2.0
168
+ vespo_lambda_neg: 2.0
169
+ vespo_lambda_pos: 3.0
170
+ vllm_enable_sleep_mode: false
171
+ vllm_gpu_memory_utilization: 0.45
172
+ vllm_group_port: 51216
173
+ vllm_importance_sampling_cap: null
174
+ vllm_importance_sampling_clip_max: 3.0
175
+ vllm_importance_sampling_clip_min: null
176
+ vllm_importance_sampling_correction: false
177
+ vllm_importance_sampling_mode: sequence_mask
178
+ vllm_max_model_length: null
179
+ vllm_mode: colocate
180
+ vllm_model_impl: vllm
181
+ vllm_server_base_url: null
182
+ vllm_server_host: 0.0.0.0
183
+ vllm_server_port: 8000
184
+ vllm_server_timeout: 240.0
185
+ vllm_structured_outputs_regex: null
186
+ vllm_tensor_parallel_size: 1
187
+ warmup_steps: 10
188
+ weight_decay: 0.0
runs/grpo_C/lexsi_provenance.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "lexsi.provenance/1",
3
+ "library": "agenttune",
4
+ "version": "1.1.3",
5
+ "git_sha": null,
6
+ "created_at": "2026-11-09T06:03:51Z",
7
+ "base_model": "CohereLabs/tiny-aya-global",
8
+ "method": "agentic.grpo",
9
+ "inputs": [],
10
+ "params": {}
11
+ }
runs/grpo_C/training_stats.json ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "PeftModelForCausalLM(\n (base_model): LoraModel(\n (model): Cohere2ForCausalLM(\n (model): Cohere2Model(\n (embed_tokens): Embedding(262144, 2048, padding_idx=0)\n (layers): ModuleList(\n (0-35): 36 x Cohere2DecoderLayer(\n (self_attn): Cohere2Attention(\n (q_proj): lora.Linear(\n (base_layer): Linear(in_features=2048, out_features=2048, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=2048, out_features=16, bias=False)\n (ref): Linear(in_features=2048, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=2048, bias=False)\n (ref): Linear(in_features=16, out_features=2048, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n (k_proj): lora.Linear(\n (base_layer): Linear(in_features=2048, out_features=512, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=2048, out_features=16, bias=False)\n (ref): Linear(in_features=2048, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=512, bias=False)\n (ref): Linear(in_features=16, out_features=512, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n (v_proj): lora.Linear(\n (base_layer): Linear(in_features=2048, out_features=512, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=2048, out_features=16, bias=False)\n (ref): Linear(in_features=2048, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=512, bias=False)\n (ref): Linear(in_features=16, out_features=512, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n (o_proj): lora.Linear(\n (base_layer): Linear(in_features=2048, out_features=2048, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=2048, out_features=16, bias=False)\n (ref): Linear(in_features=2048, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=2048, bias=False)\n (ref): Linear(in_features=16, out_features=2048, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n )\n (mlp): Cohere2MLP(\n (gate_proj): lora.Linear(\n (base_layer): Linear(in_features=2048, out_features=11008, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=2048, out_features=16, bias=False)\n (ref): Linear(in_features=2048, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=11008, bias=False)\n (ref): Linear(in_features=16, out_features=11008, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n (up_proj): lora.Linear(\n (base_layer): Linear(in_features=2048, out_features=11008, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=2048, out_features=16, bias=False)\n (ref): Linear(in_features=2048, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=11008, bias=False)\n (ref): Linear(in_features=16, out_features=11008, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n (down_proj): lora.Linear(\n (base_layer): Linear(in_features=11008, out_features=2048, bias=False)\n (lora_dropout): ModuleDict(\n (default): Dropout(p=0.05, inplace=False)\n (ref): Dropout(p=0.05, inplace=False)\n )\n (lora_A): ModuleDict(\n (default): Linear(in_features=11008, out_features=16, bias=False)\n (ref): Linear(in_features=11008, out_features=16, bias=False)\n )\n (lora_B): ModuleDict(\n (default): Linear(in_features=16, out_features=2048, bias=False)\n (ref): Linear(in_features=16, out_features=2048, bias=False)\n )\n (lora_embedding_A): ParameterDict()\n (lora_embedding_B): ParameterDict()\n (lora_magnitude_vector): ModuleDict()\n )\n (act_fn): SiLUActivation()\n )\n (input_layernorm): Cohere2LayerNorm()\n )\n )\n (norm): Cohere2LayerNorm()\n (rotary_emb): Cohere2RotaryEmbedding()\n )\n (lm_head): Linear(in_features=2048, out_features=262144, bias=False)\n )\n )\n)",
3
+ "dataset_name": "unknown",
4
+ "output_dir": "runs/grpo_C",
5
+ "agentic_mode": true,
6
+ "tools": [
7
+ "run_sql"
8
+ ],
9
+ "train_size": 445,
10
+ "eval_size": 0,
11
+ "final_loss": 8.981653369157053e-05,
12
+ "total_steps": 60,
13
+ "training_history": [],
14
+ "metrics": {
15
+ "train_runtime": 12363.7163,
16
+ "train_samples_per_second": 0.116,
17
+ "train_steps_per_second": 0.005,
18
+ "total_flos": 0.0,
19
+ "train_loss": 8.981653369157053e-05,
20
+ "epoch": 0.5405405405405406
21
+ },
22
+ "config_kwargs": {
23
+ "model": "<callable: PeftModelForCausalLM>",
24
+ "reward_funcs": [
25
+ "<function _make.<locals>.reward at 0x71a093cdb6a0>"
26
+ ],
27
+ "tools": [
28
+ "<function run_sql at 0x71a093cdac00>"
29
+ ],
30
+ "train_dataset": "Dataset({\n features: ['prompt', 'dbs', 'gold', 'ordered', 'id', 'family'],\n num_rows: 445\n})",
31
+ "output_dir": "runs/grpo_C",
32
+ "remove_unused_columns": false,
33
+ "max_steps": 60,
34
+ "num_generations": 6,
35
+ "per_device_train_batch_size": 1,
36
+ "gradient_accumulation_steps": 24,
37
+ "learning_rate": 1e-05,
38
+ "beta": 0.02,
39
+ "temperature": 1.0,
40
+ "max_completion_length": 2048,
41
+ "max_steps_per_turn": 4,
42
+ "gradient_checkpointing": true,
43
+ "bf16": true,
44
+ "logging_steps": 1,
45
+ "save_steps": 20,
46
+ "report_to": [],
47
+ "seed": 0,
48
+ "use_vllm": true,
49
+ "vllm_mode": "colocate",
50
+ "vllm_gpu_memory_utilization": 0.45,
51
+ "vllm_importance_sampling_correction": false
52
+ }
53
+ }
runs/grpo_C_rewards.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
test/dbs.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dbe47f3c9520c51417a280e71c067eb19a59f885095fc66666c41ab6c12b4e8d
3
  size 192497
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f6972585de34a2e9a2e0e5284d593d8922305442f9c087fd5056e6451cd69b1
3
  size 192497
train/dbs.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5c2a82764db7367e8570a64e50110dec46f87f775a3c953cc104d84cf5b58493
3
  size 14682025
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7b7f2dcc0271414214dca13741a42fa2cc449c35e75ad4caecd14c493250d03
3
  size 14682025