LogicPalette commited on
Commit
9223fc6
·
verified ·
1 Parent(s): 168a6b3

data update

Browse files
README.md CHANGED
@@ -35,11 +35,13 @@ for split in ("audit", "train", "test"):
35
  | `train/grpo_items.jsonl` | GRPO prompts (SFT pass rate in (0,1); 3 templates held out of SFT) |
36
  | `sft.jsonl`, `sft_all_templates.jsonl` | Scripted oracle trajectories for the SFT warm start |
37
  | `results/e0_*.json`, `results/e1_*.json`, `figure*.png` | Reward audit (FP/FN per reward) and blind rate vs DB size |
38
- | `results/traj_<model>.jsonl` | Test trajectories: base, sft, sft_ho, GRPO B, GRPO C, Command A |
39
  | `results/e4_eval.json` | Verified / single-DB accuracy, silent error, smuggling, pass^k, McNemar, bootstrap |
40
  | `results/a2_spider_lucky.json` | E6: Spider-dev lucky passes of 8 published systems (question indices only; predictions from bossben/ESMp, DBs from Zhong et al. 2020) |
41
  | `results/a3_skyrl_audit.json` | E7: SkyRL-SQL reward audit on the E0 candidates |
42
  | `results/a4_selfconsistency.json` | E8: gold-free majority vote and shaky-answer flag on the test trajectories |
 
 
43
  | `runs/grpo_<B,C>_rewards.jsonl` | Every GRPO rollout: final SQL, reward, per-DB pass vector |
44
  | `runs/grpo_<B,C>/` | `lexsi_provenance.json`, training config and stats |
45
 
 
35
  | `train/grpo_items.jsonl` | GRPO prompts (SFT pass rate in (0,1); 3 templates held out of SFT) |
36
  | `sft.jsonl`, `sft_all_templates.jsonl` | Scripted oracle trajectories for the SFT warm start |
37
  | `results/e0_*.json`, `results/e1_*.json`, `figure*.png` | Reward audit (FP/FN per reward) and blind rate vs DB size |
38
+ | `results/traj_<model>.jsonl` | Test trajectories: base, sft, sft_ho, GRPO B, GRPO C, Command A, sft_distil (E9) |
39
  | `results/e4_eval.json` | Verified / single-DB accuracy, silent error, smuggling, pass^k, McNemar, bootstrap |
40
  | `results/a2_spider_lucky.json` | E6: Spider-dev lucky passes of 8 published systems (question indices only; predictions from bossben/ESMp, DBs from Zhong et al. 2020) |
41
  | `results/a3_skyrl_audit.json` | E7: SkyRL-SQL reward audit on the E0 candidates |
42
  | `results/a4_selfconsistency.json` | E8: gold-free majority vote and shaky-answer flag on the test trajectories |
43
+ | `results/teacher_command_a.jsonl`, `results/d1_teacher_filter.json` | E9: Command A plan + SQL on the 353 SFT train items; lucky answers a single-DB filter would keep |
44
+ | `sft_distil.jsonl` | E9: the SFT trajectories with the 211 counterfactual-verified teacher endings swapped in (`traj_sft_distil` is the resulting model) |
45
  | `runs/grpo_<B,C>_rewards.jsonl` | Every GRPO rollout: final SQL, reward, per-DB pass vector |
46
  | `runs/grpo_<B,C>/` | `lexsi_provenance.json`, training config and stats |
47
 
audit/dbs.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9dc0c69f780053cdbe844ecd081c88f4b54705da9abcc574106e1c00bb794c33
3
  size 23599633
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e090035ac9993e2e2c0a37520c02708cce245245f10b1230df436ecf97ce8a15
3
  size 23599633
results/d1_teacher_filter.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "what": "D1: Command A teacher answers (plan + SQL, T=0.3) on the 353 sft_ho train items, graded on the original train DB and its train counterfactual copies",
3
+ "n": 353,
4
+ "has_sql": [
5
+ 1.0,
6
+ 0.989234439034014,
7
+ 1.0
8
+ ],
9
+ "kept_by_single_db_filter": [
10
+ 0.7308781869688386,
11
+ 0.6823088961285552,
12
+ 0.7744764114140633
13
+ ],
14
+ "kept_by_counterfactual_filter": [
15
+ 0.5977337110481586,
16
+ 0.5457927922714795,
17
+ 0.6475703133753957
18
+ ],
19
+ "lucky_kept_only_by_single_db": 47,
20
+ "lucky_share_of_single_db_kept": 0.1821705426356589,
21
+ "smuggled": 0,
22
+ "by_family_n_single_verified": {
23
+ "F1": [
24
+ 152,
25
+ 111,
26
+ 88
27
+ ],
28
+ "F2": [
29
+ 81,
30
+ 27,
31
+ 3
32
+ ],
33
+ "F3": [
34
+ 120,
35
+ 120,
36
+ 120
37
+ ]
38
+ },
39
+ "lucky_ids": [
40
+ "F1_customer_list|Delhi|apparel|S_s103",
41
+ "F1_customer_list|Delhi|home|S_s101",
42
+ "F1_customer_list|Delhi|home|S_s102",
43
+ "F1_customer_list|Indore|apparel|S_s103",
44
+ "F1_customer_list|Indore|books|S_s101",
45
+ "F1_customer_list|Indore|electronics|S_s103",
46
+ "F1_customer_list|Jaipur|books|S_s103",
47
+ "F1_customer_list|Jaipur|electronics|S_s101",
48
+ "F1_customer_list|Jaipur|home|S_s101",
49
+ "F1_customer_list|Kochi|apparel|S_s101",
50
+ "F1_customer_list|Kochi|toys|S_s103",
51
+ "F1_customer_list|Mumbai|electronics|S_s101",
52
+ "F1_customer_list|Mumbai|toys|S_s101",
53
+ "F1_customer_list|Nagpur|electronics|S_s103",
54
+ "F1_customer_list|Nagpur|grocery|S_s101",
55
+ "F1_customer_list|Nagpur|toys|S_s101",
56
+ "F1_customer_list|Pune|apparel|S_s103",
57
+ "F1_customer_list|Pune|electronics|S_s103",
58
+ "F1_customer_list|Pune|home|S_s101",
59
+ "F1_customer_list|Pune|toys|S_s101",
60
+ "F1_customer_list|Surat|apparel|S_s103",
61
+ "F1_customer_list|Surat|books|S_s102",
62
+ "F1_customer_list|Surat|electronics|S_s101",
63
+ "F2_item_value_paid|Delhi|2024|S_s103",
64
+ "F2_item_value_paid|Delhi|2025|S_s102",
65
+ "F2_item_value_paid|Delhi|2025|S_s103",
66
+ "F2_item_value_paid|Indore|2024|S_s103",
67
+ "F2_item_value_paid|Indore|2025|S_s101",
68
+ "F2_item_value_paid|Jaipur|2024|S_s102",
69
+ "F2_item_value_paid|Kochi|2024|S_s101",
70
+ "F2_item_value_paid|Mumbai|2024|S_s102",
71
+ "F2_item_value_paid|Mumbai|2025|S_s102",
72
+ "F2_item_value_paid|Nagpur|2024|S_s103",
73
+ "F2_item_value_paid|Nagpur|2025|S_s102",
74
+ "F2_item_value_paid|Surat|2024|S_s103",
75
+ "F2_item_value_paid|Surat|2025|S_s103",
76
+ "F2_paid_for_category|books|2024|S_s101",
77
+ "F2_paid_for_category|books|2024|S_s102",
78
+ "F2_paid_for_category|books|2025|S_s101",
79
+ "F2_paid_for_category|grocery|2024|S_s103",
80
+ "F2_paid_for_category|grocery|2025|S_s103",
81
+ "F2_paid_for_category|home|2024|S_s101",
82
+ "F2_paid_for_category|home|2024|S_s102",
83
+ "F2_paid_for_category|home|2025|S_s101",
84
+ "F2_paid_for_category|home|2025|S_s102",
85
+ "F2_paid_for_category|toys|2024|S_s101",
86
+ "F2_paid_for_category|toys|2025|S_s101"
87
+ ]
88
+ }
results/e4_eval.json CHANGED
@@ -292,6 +292,65 @@
292
  "4": 0.04574881312041433
293
  }
294
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
295
  "sft_ho": {
296
  "items": 331,
297
  "samples": 2648,
@@ -486,6 +545,36 @@
486
  0.6204682779456193,
487
  0.6884441087613293
488
  ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
489
  }
490
  }
491
  }
 
292
  "4": 0.04574881312041433
293
  }
294
  },
295
+ "sft_distil": {
296
+ "items": 331,
297
+ "samples": 2648,
298
+ "verified": [
299
+ 0.032854984894259816,
300
+ 0.026713379938317596,
301
+ 0.04035005250802459
302
+ ],
303
+ "single": [
304
+ 0.11216012084592145,
305
+ 0.10069811807301167,
306
+ 0.12474581492811478
307
+ ],
308
+ "silent_error": [
309
+ 0.7624622356495468,
310
+ 0.7458796916194327,
311
+ 0.7782843459309249
312
+ ],
313
+ "smuggled": [
314
+ 0.0007552870090634441,
315
+ 0.00020714704548820194,
316
+ 0.002749892319155599
317
+ ],
318
+ "has_sql": [
319
+ 0.9845166163141994,
320
+ 0.979063359109742,
321
+ 0.9885660799940115
322
+ ],
323
+ "inflation_single_minus_verified": 0.07930513595166164,
324
+ "invalid_tool_call_rate": 0.13636363636363635,
325
+ "tool_calls_per_traj": 2.1351963746223563,
326
+ "verified_by_family": {
327
+ "F1": [
328
+ 0.041666666666666664,
329
+ 0.023992942786740433,
330
+ 0.07140675348685457
331
+ ],
332
+ "F2": [
333
+ 0.006944444444444444,
334
+ 0.0019064476108571395,
335
+ 0.024962922622858687
336
+ ],
337
+ "F3": [
338
+ 0.0,
339
+ 0.0,
340
+ 0.010558440816003448
341
+ ],
342
+ "F4": [
343
+ 0.04264018691588785,
344
+ 0.03404933943902742,
345
+ 0.05327900135886442
346
+ ]
347
+ },
348
+ "pass^k": {
349
+ "1": 0.032854984894259816,
350
+ "2": 0.017371601208459216,
351
+ "4": 0.01126456624946051
352
+ }
353
+ },
354
  "sft_ho": {
355
  "items": 331,
356
  "samples": 2648,
 
545
  0.6204682779456193,
546
  0.6884441087613293
547
  ]
548
+ },
549
+ "sft_distil_vs_sft_ho": {
550
+ "n_items": 331,
551
+ "mcnemar_p_sample0": 0.0009122341871261597,
552
+ "verified_diff_bootstrap95": [
553
+ -0.061555891238670696,
554
+ -0.08270392749244714,
555
+ -0.0411631419939577
556
+ ],
557
+ "silent_mcnemar_p_sample0": 0.06353719866713946,
558
+ "silent_error_diff_bootstrap95": [
559
+ 0.10800604229607251,
560
+ 0.06722054380664652,
561
+ 0.14841389728096677
562
+ ]
563
+ },
564
+ "B_vs_sft_distil": {
565
+ "n_items": 331,
566
+ "mcnemar_p_sample0": 4.405936238072172e-08,
567
+ "verified_diff_bootstrap95": [
568
+ 0.09403323262839879,
569
+ 0.06759818731117825,
570
+ 0.1216012084592145
571
+ ],
572
+ "silent_mcnemar_p_sample0": 0.000893531059458462,
573
+ "silent_error_diff_bootstrap95": [
574
+ -0.12197885196374622,
575
+ -0.16200906344410876,
576
+ -0.0823262839879154
577
+ ]
578
  }
579
  }
580
  }
results/teacher_command_a.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
results/traj_sft_distil.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
sft_distil.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:ec90760d9b784d752ec2b3b245d864c687ac32b86791b5868420c1fae1fd5299
3
  size 192497
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:86fbf57c65c65b1a96b4508cacacf8a968c00d41fd3e6e8826daeffed02ef3bc
3
  size 192497
train/dbs.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0775d3ef895c52ee0909c603f3f34a215f9319b8c5f38e0819b29967e1a24558
3
  size 14682025
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32326f7a6e7ccab602f13692c3a4f7ef56557448c71fd110027517f70c480901
3
  size 14682025