fisherman611 commited on
Commit
08d62aa
·
verified ·
1 Parent(s): fa119ae

backup part 17

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed10/best_by_benchmark.json +72 -0
  2. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed10/best_overall.json +77 -0
  3. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed10/eval_results.json +514 -0
  4. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed100/best_by_benchmark.json +72 -0
  5. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed100/best_overall.json +77 -0
  6. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed100/eval_results.json +514 -0
  7. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed1234/best_by_benchmark.json +72 -0
  8. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed1234/best_overall.json +77 -0
  9. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed1234/eval_results.json +514 -0
  10. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed42/best_by_benchmark.json +72 -0
  11. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed42/best_overall.json +77 -0
  12. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed42/eval_results.json +514 -0
  13. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed50/best_by_benchmark.json +72 -0
  14. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed50/best_overall.json +77 -0
  15. text2sql_distillation_draft/results/eval/llama_baselines/llama/seed50/eval_results.json +514 -0
  16. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log +0 -0
  17. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log +0 -0
  18. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log +0 -0
  19. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  20. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  21. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log +0 -0
  22. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log +0 -0
  23. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log +0 -0
  24. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log +0 -0
  25. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log +0 -0
  26. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  27. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  28. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log +0 -0
  29. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log +0 -0
  30. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log +0 -0
  31. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log +0 -0
  32. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log +0 -0
  33. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  34. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  35. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log +0 -0
  36. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log +0 -0
  37. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log +0 -0
  38. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log +0 -0
  39. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log +0 -0
  40. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  41. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  42. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log +0 -0
  43. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log +0 -0
  44. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log +0 -0
  45. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log +0 -0
  46. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log +0 -0
  47. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  48. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log +0 -0
  49. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log +0 -0
  50. text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log +0 -0
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed10/best_by_benchmark.json ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 10,
3
+ "generated_at": "2026-07-20T12:37:18.609572+00:00",
4
+ "primary_metric": "Execution Accuracy",
5
+ "tie_breaker": "Exact Match",
6
+ "best_by_benchmark": {
7
+ "spider_data": {
8
+ "run_name": "baselines__train_rkl__ckpt872__test",
9
+ "scores": {
10
+ "Exact Match": 0.549,
11
+ "Execution Accuracy": 0.607
12
+ },
13
+ "artifacts": {
14
+ "run_name": "baselines__train_rkl__ckpt872__test",
15
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
16
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
17
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
18
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
19
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
20
+ "ckpt_step": "872"
21
+ }
22
+ },
23
+ "spider_dk": {
24
+ "run_name": "baselines__train_csd__ckpt654__test",
25
+ "scores": {
26
+ "Exact Match": 0.407,
27
+ "Execution Accuracy": 0.464
28
+ },
29
+ "artifacts": {
30
+ "run_name": "baselines__train_csd__ckpt654__test",
31
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
32
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
33
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
34
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
35
+ "train_script": "scripts/llama/baselines/train_csd.sh",
36
+ "ckpt_step": "654"
37
+ }
38
+ },
39
+ "spider_realistic": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "scores": {
42
+ "Exact Match": 0.461,
43
+ "Execution Accuracy": 0.496
44
+ },
45
+ "artifacts": {
46
+ "run_name": "baselines__train_rkl__ckpt872__test",
47
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
48
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
49
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
50
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
51
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
52
+ "ckpt_step": "872"
53
+ }
54
+ },
55
+ "spider_syn": {
56
+ "run_name": "baselines__train_rkl__ckpt872__test",
57
+ "scores": {
58
+ "Exact Match": 0.411,
59
+ "Execution Accuracy": 0.472
60
+ },
61
+ "artifacts": {
62
+ "run_name": "baselines__train_rkl__ckpt872__test",
63
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
64
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
65
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
66
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
67
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
68
+ "ckpt_step": "872"
69
+ }
70
+ }
71
+ }
72
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed10/best_overall.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 10,
3
+ "generated_at": "2026-07-20T12:37:18.609572+00:00",
4
+ "aggregation": {
5
+ "benchmarks": [
6
+ "spider_data",
7
+ "spider_syn",
8
+ "spider_realistic",
9
+ "spider_dk"
10
+ ],
11
+ "primary_metric": "mean Execution Accuracy",
12
+ "tie_breaker": "mean Exact Match"
13
+ },
14
+ "best": {
15
+ "run_name": "baselines__train_csd__ckpt654__test",
16
+ "mean_scores": {
17
+ "Exact Match": 0.44875,
18
+ "Execution Accuracy": 0.505
19
+ },
20
+ "per_benchmark": {
21
+ "spider_data": {
22
+ "Exact Match": 0.535,
23
+ "Execution Accuracy": 0.594
24
+ },
25
+ "spider_syn": {
26
+ "Exact Match": 0.406,
27
+ "Execution Accuracy": 0.47
28
+ },
29
+ "spider_realistic": {
30
+ "Exact Match": 0.447,
31
+ "Execution Accuracy": 0.492
32
+ },
33
+ "spider_dk": {
34
+ "Exact Match": 0.407,
35
+ "Execution Accuracy": 0.464
36
+ }
37
+ },
38
+ "artifacts": {
39
+ "spider_data": {
40
+ "run_name": "baselines__train_csd__ckpt654__test",
41
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
42
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
43
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
44
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
45
+ "train_script": "scripts/llama/baselines/train_csd.sh",
46
+ "ckpt_step": "654"
47
+ },
48
+ "spider_syn": {
49
+ "run_name": "baselines__train_csd__ckpt654__test",
50
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
51
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
52
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
53
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
54
+ "train_script": "scripts/llama/baselines/train_csd.sh",
55
+ "ckpt_step": "654"
56
+ },
57
+ "spider_realistic": {
58
+ "run_name": "baselines__train_csd__ckpt654__test",
59
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
60
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
61
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
62
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
63
+ "train_script": "scripts/llama/baselines/train_csd.sh",
64
+ "ckpt_step": "654"
65
+ },
66
+ "spider_dk": {
67
+ "run_name": "baselines__train_csd__ckpt654__test",
68
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
69
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
70
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
71
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
72
+ "train_script": "scripts/llama/baselines/train_csd.sh",
73
+ "ckpt_step": "654"
74
+ }
75
+ }
76
+ }
77
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed10/eval_results.json ADDED
@@ -0,0 +1,514 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 10,
3
+ "generated_at": "2026-07-20T12:37:18.609572+00:00",
4
+ "selection": {
5
+ "primary_metric": "Execution Accuracy",
6
+ "tie_breaker": "Exact Match"
7
+ },
8
+ "results": [
9
+ {
10
+ "run_name": "baselines__train_amid__ckpt872__test",
11
+ "benchmark": "spider_data",
12
+ "seed": 10,
13
+ "scores": {
14
+ "Exact Match": 0.506,
15
+ "Execution Accuracy": 0.576
16
+ },
17
+ "artifacts": {
18
+ "run_name": "baselines__train_amid__ckpt872__test",
19
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
20
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
21
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
22
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
23
+ "train_script": "scripts/llama/baselines/train_amid.sh",
24
+ "ckpt_step": "872"
25
+ }
26
+ },
27
+ {
28
+ "run_name": "baselines__train_amid__ckpt872__test",
29
+ "benchmark": "spider_dk",
30
+ "seed": 10,
31
+ "scores": {
32
+ "Exact Match": 0.381,
33
+ "Execution Accuracy": 0.454
34
+ },
35
+ "artifacts": {
36
+ "run_name": "baselines__train_amid__ckpt872__test",
37
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
38
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
39
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
40
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
41
+ "train_script": "scripts/llama/baselines/train_amid.sh",
42
+ "ckpt_step": "872"
43
+ }
44
+ },
45
+ {
46
+ "run_name": "baselines__train_amid__ckpt872__test",
47
+ "benchmark": "spider_realistic",
48
+ "seed": 10,
49
+ "scores": {
50
+ "Exact Match": 0.417,
51
+ "Execution Accuracy": 0.461
52
+ },
53
+ "artifacts": {
54
+ "run_name": "baselines__train_amid__ckpt872__test",
55
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
56
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
57
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
58
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
59
+ "train_script": "scripts/llama/baselines/train_amid.sh",
60
+ "ckpt_step": "872"
61
+ }
62
+ },
63
+ {
64
+ "run_name": "baselines__train_amid__ckpt872__test",
65
+ "benchmark": "spider_syn",
66
+ "seed": 10,
67
+ "scores": {
68
+ "Exact Match": 0.361,
69
+ "Execution Accuracy": 0.425
70
+ },
71
+ "artifacts": {
72
+ "run_name": "baselines__train_amid__ckpt872__test",
73
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
74
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
75
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
76
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
77
+ "train_script": "scripts/llama/baselines/train_amid.sh",
78
+ "ckpt_step": "872"
79
+ }
80
+ },
81
+ {
82
+ "run_name": "baselines__train_csd__ckpt654__test",
83
+ "benchmark": "spider_data",
84
+ "seed": 10,
85
+ "scores": {
86
+ "Exact Match": 0.535,
87
+ "Execution Accuracy": 0.594
88
+ },
89
+ "artifacts": {
90
+ "run_name": "baselines__train_csd__ckpt654__test",
91
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
92
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
93
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
94
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
95
+ "train_script": "scripts/llama/baselines/train_csd.sh",
96
+ "ckpt_step": "654"
97
+ }
98
+ },
99
+ {
100
+ "run_name": "baselines__train_csd__ckpt654__test",
101
+ "benchmark": "spider_dk",
102
+ "seed": 10,
103
+ "scores": {
104
+ "Exact Match": 0.407,
105
+ "Execution Accuracy": 0.464
106
+ },
107
+ "artifacts": {
108
+ "run_name": "baselines__train_csd__ckpt654__test",
109
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
110
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
111
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
112
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
113
+ "train_script": "scripts/llama/baselines/train_csd.sh",
114
+ "ckpt_step": "654"
115
+ }
116
+ },
117
+ {
118
+ "run_name": "baselines__train_csd__ckpt654__test",
119
+ "benchmark": "spider_realistic",
120
+ "seed": 10,
121
+ "scores": {
122
+ "Exact Match": 0.447,
123
+ "Execution Accuracy": 0.492
124
+ },
125
+ "artifacts": {
126
+ "run_name": "baselines__train_csd__ckpt654__test",
127
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
128
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
129
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
130
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
131
+ "train_script": "scripts/llama/baselines/train_csd.sh",
132
+ "ckpt_step": "654"
133
+ }
134
+ },
135
+ {
136
+ "run_name": "baselines__train_csd__ckpt654__test",
137
+ "benchmark": "spider_syn",
138
+ "seed": 10,
139
+ "scores": {
140
+ "Exact Match": 0.406,
141
+ "Execution Accuracy": 0.47
142
+ },
143
+ "artifacts": {
144
+ "run_name": "baselines__train_csd__ckpt654__test",
145
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
146
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
147
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
148
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
149
+ "train_script": "scripts/llama/baselines/train_csd.sh",
150
+ "ckpt_step": "654"
151
+ }
152
+ },
153
+ {
154
+ "run_name": "baselines__train_distillm__ckpt654__test",
155
+ "benchmark": "spider_data",
156
+ "seed": 10,
157
+ "scores": {
158
+ "Exact Match": 0.505,
159
+ "Execution Accuracy": 0.585
160
+ },
161
+ "artifacts": {
162
+ "run_name": "baselines__train_distillm__ckpt654__test",
163
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
164
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
165
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
166
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
167
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
168
+ "ckpt_step": "654"
169
+ }
170
+ },
171
+ {
172
+ "run_name": "baselines__train_distillm__ckpt654__test",
173
+ "benchmark": "spider_dk",
174
+ "seed": 10,
175
+ "scores": {
176
+ "Exact Match": 0.387,
177
+ "Execution Accuracy": 0.443
178
+ },
179
+ "artifacts": {
180
+ "run_name": "baselines__train_distillm__ckpt654__test",
181
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
182
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
183
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
184
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
185
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
186
+ "ckpt_step": "654"
187
+ }
188
+ },
189
+ {
190
+ "run_name": "baselines__train_distillm__ckpt654__test",
191
+ "benchmark": "spider_realistic",
192
+ "seed": 10,
193
+ "scores": {
194
+ "Exact Match": 0.419,
195
+ "Execution Accuracy": 0.467
196
+ },
197
+ "artifacts": {
198
+ "run_name": "baselines__train_distillm__ckpt654__test",
199
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
200
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
201
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
202
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
203
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
204
+ "ckpt_step": "654"
205
+ }
206
+ },
207
+ {
208
+ "run_name": "baselines__train_distillm__ckpt654__test",
209
+ "benchmark": "spider_syn",
210
+ "seed": 10,
211
+ "scores": {
212
+ "Exact Match": 0.372,
213
+ "Execution Accuracy": 0.442
214
+ },
215
+ "artifacts": {
216
+ "run_name": "baselines__train_distillm__ckpt654__test",
217
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
218
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
219
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
220
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
221
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
222
+ "ckpt_step": "654"
223
+ }
224
+ },
225
+ {
226
+ "run_name": "baselines__train_fkl__ckpt872__test",
227
+ "benchmark": "spider_data",
228
+ "seed": 10,
229
+ "scores": {
230
+ "Exact Match": 0.506,
231
+ "Execution Accuracy": 0.576
232
+ },
233
+ "artifacts": {
234
+ "run_name": "baselines__train_fkl__ckpt872__test",
235
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
236
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
237
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
238
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
239
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
240
+ "ckpt_step": "872"
241
+ }
242
+ },
243
+ {
244
+ "run_name": "baselines__train_fkl__ckpt872__test",
245
+ "benchmark": "spider_dk",
246
+ "seed": 10,
247
+ "scores": {
248
+ "Exact Match": 0.368,
249
+ "Execution Accuracy": 0.439
250
+ },
251
+ "artifacts": {
252
+ "run_name": "baselines__train_fkl__ckpt872__test",
253
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
254
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
255
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
256
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
257
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
258
+ "ckpt_step": "872"
259
+ }
260
+ },
261
+ {
262
+ "run_name": "baselines__train_fkl__ckpt872__test",
263
+ "benchmark": "spider_realistic",
264
+ "seed": 10,
265
+ "scores": {
266
+ "Exact Match": 0.39,
267
+ "Execution Accuracy": 0.449
268
+ },
269
+ "artifacts": {
270
+ "run_name": "baselines__train_fkl__ckpt872__test",
271
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
272
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
273
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
274
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
275
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
276
+ "ckpt_step": "872"
277
+ }
278
+ },
279
+ {
280
+ "run_name": "baselines__train_fkl__ckpt872__test",
281
+ "benchmark": "spider_syn",
282
+ "seed": 10,
283
+ "scores": {
284
+ "Exact Match": 0.365,
285
+ "Execution Accuracy": 0.437
286
+ },
287
+ "artifacts": {
288
+ "run_name": "baselines__train_fkl__ckpt872__test",
289
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
290
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
291
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
292
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
293
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
294
+ "ckpt_step": "872"
295
+ }
296
+ },
297
+ {
298
+ "run_name": "baselines__train_rkl__ckpt872__test",
299
+ "benchmark": "spider_data",
300
+ "seed": 10,
301
+ "scores": {
302
+ "Exact Match": 0.549,
303
+ "Execution Accuracy": 0.607
304
+ },
305
+ "artifacts": {
306
+ "run_name": "baselines__train_rkl__ckpt872__test",
307
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
308
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
309
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
310
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
311
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
312
+ "ckpt_step": "872"
313
+ }
314
+ },
315
+ {
316
+ "run_name": "baselines__train_rkl__ckpt872__test",
317
+ "benchmark": "spider_dk",
318
+ "seed": 10,
319
+ "scores": {
320
+ "Exact Match": 0.393,
321
+ "Execution Accuracy": 0.439
322
+ },
323
+ "artifacts": {
324
+ "run_name": "baselines__train_rkl__ckpt872__test",
325
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
326
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
327
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
328
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
329
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
330
+ "ckpt_step": "872"
331
+ }
332
+ },
333
+ {
334
+ "run_name": "baselines__train_rkl__ckpt872__test",
335
+ "benchmark": "spider_realistic",
336
+ "seed": 10,
337
+ "scores": {
338
+ "Exact Match": 0.461,
339
+ "Execution Accuracy": 0.496
340
+ },
341
+ "artifacts": {
342
+ "run_name": "baselines__train_rkl__ckpt872__test",
343
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
344
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
345
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
346
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
347
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
348
+ "ckpt_step": "872"
349
+ }
350
+ },
351
+ {
352
+ "run_name": "baselines__train_rkl__ckpt872__test",
353
+ "benchmark": "spider_syn",
354
+ "seed": 10,
355
+ "scores": {
356
+ "Exact Match": 0.411,
357
+ "Execution Accuracy": 0.472
358
+ },
359
+ "artifacts": {
360
+ "run_name": "baselines__train_rkl__ckpt872__test",
361
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
362
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
363
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
364
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
365
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
366
+ "ckpt_step": "872"
367
+ }
368
+ },
369
+ {
370
+ "run_name": "baselines__train_sfkl__ckpt654__test",
371
+ "benchmark": "spider_data",
372
+ "seed": 10,
373
+ "scores": {
374
+ "Exact Match": 0.499,
375
+ "Execution Accuracy": 0.574
376
+ },
377
+ "artifacts": {
378
+ "run_name": "baselines__train_sfkl__ckpt654__test",
379
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
380
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
381
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
382
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
383
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
384
+ "ckpt_step": "654"
385
+ }
386
+ },
387
+ {
388
+ "run_name": "baselines__train_sfkl__ckpt654__test",
389
+ "benchmark": "spider_dk",
390
+ "seed": 10,
391
+ "scores": {
392
+ "Exact Match": 0.391,
393
+ "Execution Accuracy": 0.458
394
+ },
395
+ "artifacts": {
396
+ "run_name": "baselines__train_sfkl__ckpt654__test",
397
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
398
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
399
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
400
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
401
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
402
+ "ckpt_step": "654"
403
+ }
404
+ },
405
+ {
406
+ "run_name": "baselines__train_sfkl__ckpt654__test",
407
+ "benchmark": "spider_realistic",
408
+ "seed": 10,
409
+ "scores": {
410
+ "Exact Match": 0.437,
411
+ "Execution Accuracy": 0.482
412
+ },
413
+ "artifacts": {
414
+ "run_name": "baselines__train_sfkl__ckpt654__test",
415
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
416
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
417
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
418
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
419
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
420
+ "ckpt_step": "654"
421
+ }
422
+ },
423
+ {
424
+ "run_name": "baselines__train_sfkl__ckpt654__test",
425
+ "benchmark": "spider_syn",
426
+ "seed": 10,
427
+ "scores": {
428
+ "Exact Match": 0.381,
429
+ "Execution Accuracy": 0.428
430
+ },
431
+ "artifacts": {
432
+ "run_name": "baselines__train_sfkl__ckpt654__test",
433
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
434
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
435
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
436
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
437
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
438
+ "ckpt_step": "654"
439
+ }
440
+ },
441
+ {
442
+ "run_name": "baselines__train_srkl__ckpt436__test",
443
+ "benchmark": "spider_data",
444
+ "seed": 10,
445
+ "scores": {
446
+ "Exact Match": 0.475,
447
+ "Execution Accuracy": 0.551
448
+ },
449
+ "artifacts": {
450
+ "run_name": "baselines__train_srkl__ckpt436__test",
451
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
452
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed10/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
453
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
454
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
455
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
456
+ "ckpt_step": "436"
457
+ }
458
+ },
459
+ {
460
+ "run_name": "baselines__train_srkl__ckpt436__test",
461
+ "benchmark": "spider_dk",
462
+ "seed": 10,
463
+ "scores": {
464
+ "Exact Match": 0.346,
465
+ "Execution Accuracy": 0.407
466
+ },
467
+ "artifacts": {
468
+ "run_name": "baselines__train_srkl__ckpt436__test",
469
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
470
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed10/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
471
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed10/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
472
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
473
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
474
+ "ckpt_step": "436"
475
+ }
476
+ },
477
+ {
478
+ "run_name": "baselines__train_srkl__ckpt436__test",
479
+ "benchmark": "spider_realistic",
480
+ "seed": 10,
481
+ "scores": {
482
+ "Exact Match": 0.386,
483
+ "Execution Accuracy": 0.439
484
+ },
485
+ "artifacts": {
486
+ "run_name": "baselines__train_srkl__ckpt436__test",
487
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
488
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed10/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
489
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed10/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
490
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
491
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
492
+ "ckpt_step": "436"
493
+ }
494
+ },
495
+ {
496
+ "run_name": "baselines__train_srkl__ckpt436__test",
497
+ "benchmark": "spider_syn",
498
+ "seed": 10,
499
+ "scores": {
500
+ "Exact Match": 0.356,
501
+ "Execution Accuracy": 0.409
502
+ },
503
+ "artifacts": {
504
+ "run_name": "baselines__train_srkl__ckpt436__test",
505
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
506
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed10/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
507
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed10/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
508
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
509
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
510
+ "ckpt_step": "436"
511
+ }
512
+ }
513
+ ]
514
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed100/best_by_benchmark.json ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 100,
3
+ "generated_at": "2026-07-20T12:37:18.613555+00:00",
4
+ "primary_metric": "Execution Accuracy",
5
+ "tie_breaker": "Exact Match",
6
+ "best_by_benchmark": {
7
+ "spider_data": {
8
+ "run_name": "baselines__train_rkl__ckpt872__test",
9
+ "scores": {
10
+ "Exact Match": 0.553,
11
+ "Execution Accuracy": 0.617
12
+ },
13
+ "artifacts": {
14
+ "run_name": "baselines__train_rkl__ckpt872__test",
15
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
16
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
17
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
18
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
19
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
20
+ "ckpt_step": "872"
21
+ }
22
+ },
23
+ "spider_dk": {
24
+ "run_name": "baselines__train_csd__ckpt654__test",
25
+ "scores": {
26
+ "Exact Match": 0.43,
27
+ "Execution Accuracy": 0.486
28
+ },
29
+ "artifacts": {
30
+ "run_name": "baselines__train_csd__ckpt654__test",
31
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
32
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
33
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
34
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
35
+ "train_script": "scripts/llama/baselines/train_csd.sh",
36
+ "ckpt_step": "654"
37
+ }
38
+ },
39
+ "spider_realistic": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "scores": {
42
+ "Exact Match": 0.449,
43
+ "Execution Accuracy": 0.502
44
+ },
45
+ "artifacts": {
46
+ "run_name": "baselines__train_rkl__ckpt872__test",
47
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
48
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
49
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
50
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
51
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
52
+ "ckpt_step": "872"
53
+ }
54
+ },
55
+ "spider_syn": {
56
+ "run_name": "baselines__train_rkl__ckpt872__test",
57
+ "scores": {
58
+ "Exact Match": 0.41,
59
+ "Execution Accuracy": 0.479
60
+ },
61
+ "artifacts": {
62
+ "run_name": "baselines__train_rkl__ckpt872__test",
63
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
64
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
65
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
66
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
67
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
68
+ "ckpt_step": "872"
69
+ }
70
+ }
71
+ }
72
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed100/best_overall.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 100,
3
+ "generated_at": "2026-07-20T12:37:18.613555+00:00",
4
+ "aggregation": {
5
+ "benchmarks": [
6
+ "spider_data",
7
+ "spider_syn",
8
+ "spider_realistic",
9
+ "spider_dk"
10
+ ],
11
+ "primary_metric": "mean Execution Accuracy",
12
+ "tie_breaker": "mean Exact Match"
13
+ },
14
+ "best": {
15
+ "run_name": "baselines__train_csd__ckpt654__test",
16
+ "mean_scores": {
17
+ "Exact Match": 0.45525,
18
+ "Execution Accuracy": 0.509
19
+ },
20
+ "per_benchmark": {
21
+ "spider_data": {
22
+ "Exact Match": 0.537,
23
+ "Execution Accuracy": 0.596
24
+ },
25
+ "spider_syn": {
26
+ "Exact Match": 0.407,
27
+ "Execution Accuracy": 0.464
28
+ },
29
+ "spider_realistic": {
30
+ "Exact Match": 0.447,
31
+ "Execution Accuracy": 0.49
32
+ },
33
+ "spider_dk": {
34
+ "Exact Match": 0.43,
35
+ "Execution Accuracy": 0.486
36
+ }
37
+ },
38
+ "artifacts": {
39
+ "spider_data": {
40
+ "run_name": "baselines__train_csd__ckpt654__test",
41
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
42
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
43
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
44
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
45
+ "train_script": "scripts/llama/baselines/train_csd.sh",
46
+ "ckpt_step": "654"
47
+ },
48
+ "spider_syn": {
49
+ "run_name": "baselines__train_csd__ckpt654__test",
50
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
51
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
52
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
53
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
54
+ "train_script": "scripts/llama/baselines/train_csd.sh",
55
+ "ckpt_step": "654"
56
+ },
57
+ "spider_realistic": {
58
+ "run_name": "baselines__train_csd__ckpt654__test",
59
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
60
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
61
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
62
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
63
+ "train_script": "scripts/llama/baselines/train_csd.sh",
64
+ "ckpt_step": "654"
65
+ },
66
+ "spider_dk": {
67
+ "run_name": "baselines__train_csd__ckpt654__test",
68
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
69
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
70
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
71
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
72
+ "train_script": "scripts/llama/baselines/train_csd.sh",
73
+ "ckpt_step": "654"
74
+ }
75
+ }
76
+ }
77
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed100/eval_results.json ADDED
@@ -0,0 +1,514 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 100,
3
+ "generated_at": "2026-07-20T12:37:18.613555+00:00",
4
+ "selection": {
5
+ "primary_metric": "Execution Accuracy",
6
+ "tie_breaker": "Exact Match"
7
+ },
8
+ "results": [
9
+ {
10
+ "run_name": "baselines__train_amid__ckpt872__test",
11
+ "benchmark": "spider_data",
12
+ "seed": 100,
13
+ "scores": {
14
+ "Exact Match": 0.51,
15
+ "Execution Accuracy": 0.577
16
+ },
17
+ "artifacts": {
18
+ "run_name": "baselines__train_amid__ckpt872__test",
19
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
20
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
21
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
22
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
23
+ "train_script": "scripts/llama/baselines/train_amid.sh",
24
+ "ckpt_step": "872"
25
+ }
26
+ },
27
+ {
28
+ "run_name": "baselines__train_amid__ckpt872__test",
29
+ "benchmark": "spider_dk",
30
+ "seed": 100,
31
+ "scores": {
32
+ "Exact Match": 0.391,
33
+ "Execution Accuracy": 0.465
34
+ },
35
+ "artifacts": {
36
+ "run_name": "baselines__train_amid__ckpt872__test",
37
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
38
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
39
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
40
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
41
+ "train_script": "scripts/llama/baselines/train_amid.sh",
42
+ "ckpt_step": "872"
43
+ }
44
+ },
45
+ {
46
+ "run_name": "baselines__train_amid__ckpt872__test",
47
+ "benchmark": "spider_realistic",
48
+ "seed": 100,
49
+ "scores": {
50
+ "Exact Match": 0.407,
51
+ "Execution Accuracy": 0.459
52
+ },
53
+ "artifacts": {
54
+ "run_name": "baselines__train_amid__ckpt872__test",
55
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
56
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
57
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
58
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
59
+ "train_script": "scripts/llama/baselines/train_amid.sh",
60
+ "ckpt_step": "872"
61
+ }
62
+ },
63
+ {
64
+ "run_name": "baselines__train_amid__ckpt872__test",
65
+ "benchmark": "spider_syn",
66
+ "seed": 100,
67
+ "scores": {
68
+ "Exact Match": 0.365,
69
+ "Execution Accuracy": 0.42
70
+ },
71
+ "artifacts": {
72
+ "run_name": "baselines__train_amid__ckpt872__test",
73
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
74
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
75
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
76
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
77
+ "train_script": "scripts/llama/baselines/train_amid.sh",
78
+ "ckpt_step": "872"
79
+ }
80
+ },
81
+ {
82
+ "run_name": "baselines__train_csd__ckpt654__test",
83
+ "benchmark": "spider_data",
84
+ "seed": 100,
85
+ "scores": {
86
+ "Exact Match": 0.537,
87
+ "Execution Accuracy": 0.596
88
+ },
89
+ "artifacts": {
90
+ "run_name": "baselines__train_csd__ckpt654__test",
91
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
92
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
93
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
94
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
95
+ "train_script": "scripts/llama/baselines/train_csd.sh",
96
+ "ckpt_step": "654"
97
+ }
98
+ },
99
+ {
100
+ "run_name": "baselines__train_csd__ckpt654__test",
101
+ "benchmark": "spider_dk",
102
+ "seed": 100,
103
+ "scores": {
104
+ "Exact Match": 0.43,
105
+ "Execution Accuracy": 0.486
106
+ },
107
+ "artifacts": {
108
+ "run_name": "baselines__train_csd__ckpt654__test",
109
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
110
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
111
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
112
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
113
+ "train_script": "scripts/llama/baselines/train_csd.sh",
114
+ "ckpt_step": "654"
115
+ }
116
+ },
117
+ {
118
+ "run_name": "baselines__train_csd__ckpt654__test",
119
+ "benchmark": "spider_realistic",
120
+ "seed": 100,
121
+ "scores": {
122
+ "Exact Match": 0.447,
123
+ "Execution Accuracy": 0.49
124
+ },
125
+ "artifacts": {
126
+ "run_name": "baselines__train_csd__ckpt654__test",
127
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
128
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
129
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
130
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
131
+ "train_script": "scripts/llama/baselines/train_csd.sh",
132
+ "ckpt_step": "654"
133
+ }
134
+ },
135
+ {
136
+ "run_name": "baselines__train_csd__ckpt654__test",
137
+ "benchmark": "spider_syn",
138
+ "seed": 100,
139
+ "scores": {
140
+ "Exact Match": 0.407,
141
+ "Execution Accuracy": 0.464
142
+ },
143
+ "artifacts": {
144
+ "run_name": "baselines__train_csd__ckpt654__test",
145
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
146
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
147
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
148
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
149
+ "train_script": "scripts/llama/baselines/train_csd.sh",
150
+ "ckpt_step": "654"
151
+ }
152
+ },
153
+ {
154
+ "run_name": "baselines__train_distillm__ckpt654__test",
155
+ "benchmark": "spider_data",
156
+ "seed": 100,
157
+ "scores": {
158
+ "Exact Match": 0.506,
159
+ "Execution Accuracy": 0.585
160
+ },
161
+ "artifacts": {
162
+ "run_name": "baselines__train_distillm__ckpt654__test",
163
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
164
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
165
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
166
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
167
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
168
+ "ckpt_step": "654"
169
+ }
170
+ },
171
+ {
172
+ "run_name": "baselines__train_distillm__ckpt654__test",
173
+ "benchmark": "spider_dk",
174
+ "seed": 100,
175
+ "scores": {
176
+ "Exact Match": 0.385,
177
+ "Execution Accuracy": 0.462
178
+ },
179
+ "artifacts": {
180
+ "run_name": "baselines__train_distillm__ckpt654__test",
181
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
182
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
183
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
184
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
185
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
186
+ "ckpt_step": "654"
187
+ }
188
+ },
189
+ {
190
+ "run_name": "baselines__train_distillm__ckpt654__test",
191
+ "benchmark": "spider_realistic",
192
+ "seed": 100,
193
+ "scores": {
194
+ "Exact Match": 0.433,
195
+ "Execution Accuracy": 0.467
196
+ },
197
+ "artifacts": {
198
+ "run_name": "baselines__train_distillm__ckpt654__test",
199
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
200
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
201
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
202
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
203
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
204
+ "ckpt_step": "654"
205
+ }
206
+ },
207
+ {
208
+ "run_name": "baselines__train_distillm__ckpt654__test",
209
+ "benchmark": "spider_syn",
210
+ "seed": 100,
211
+ "scores": {
212
+ "Exact Match": 0.387,
213
+ "Execution Accuracy": 0.437
214
+ },
215
+ "artifacts": {
216
+ "run_name": "baselines__train_distillm__ckpt654__test",
217
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
218
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
219
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
220
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
221
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
222
+ "ckpt_step": "654"
223
+ }
224
+ },
225
+ {
226
+ "run_name": "baselines__train_fkl__ckpt872__test",
227
+ "benchmark": "spider_data",
228
+ "seed": 100,
229
+ "scores": {
230
+ "Exact Match": 0.517,
231
+ "Execution Accuracy": 0.591
232
+ },
233
+ "artifacts": {
234
+ "run_name": "baselines__train_fkl__ckpt872__test",
235
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
236
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
237
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
238
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
239
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
240
+ "ckpt_step": "872"
241
+ }
242
+ },
243
+ {
244
+ "run_name": "baselines__train_fkl__ckpt872__test",
245
+ "benchmark": "spider_dk",
246
+ "seed": 100,
247
+ "scores": {
248
+ "Exact Match": 0.383,
249
+ "Execution Accuracy": 0.452
250
+ },
251
+ "artifacts": {
252
+ "run_name": "baselines__train_fkl__ckpt872__test",
253
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
254
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
255
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
256
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
257
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
258
+ "ckpt_step": "872"
259
+ }
260
+ },
261
+ {
262
+ "run_name": "baselines__train_fkl__ckpt872__test",
263
+ "benchmark": "spider_realistic",
264
+ "seed": 100,
265
+ "scores": {
266
+ "Exact Match": 0.415,
267
+ "Execution Accuracy": 0.469
268
+ },
269
+ "artifacts": {
270
+ "run_name": "baselines__train_fkl__ckpt872__test",
271
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
272
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
273
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
274
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
275
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
276
+ "ckpt_step": "872"
277
+ }
278
+ },
279
+ {
280
+ "run_name": "baselines__train_fkl__ckpt872__test",
281
+ "benchmark": "spider_syn",
282
+ "seed": 100,
283
+ "scores": {
284
+ "Exact Match": 0.369,
285
+ "Execution Accuracy": 0.426
286
+ },
287
+ "artifacts": {
288
+ "run_name": "baselines__train_fkl__ckpt872__test",
289
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
290
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
291
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
292
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
293
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
294
+ "ckpt_step": "872"
295
+ }
296
+ },
297
+ {
298
+ "run_name": "baselines__train_rkl__ckpt872__test",
299
+ "benchmark": "spider_data",
300
+ "seed": 100,
301
+ "scores": {
302
+ "Exact Match": 0.553,
303
+ "Execution Accuracy": 0.617
304
+ },
305
+ "artifacts": {
306
+ "run_name": "baselines__train_rkl__ckpt872__test",
307
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
308
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
309
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
310
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
311
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
312
+ "ckpt_step": "872"
313
+ }
314
+ },
315
+ {
316
+ "run_name": "baselines__train_rkl__ckpt872__test",
317
+ "benchmark": "spider_dk",
318
+ "seed": 100,
319
+ "scores": {
320
+ "Exact Match": 0.379,
321
+ "Execution Accuracy": 0.422
322
+ },
323
+ "artifacts": {
324
+ "run_name": "baselines__train_rkl__ckpt872__test",
325
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
326
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
327
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
328
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
329
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
330
+ "ckpt_step": "872"
331
+ }
332
+ },
333
+ {
334
+ "run_name": "baselines__train_rkl__ckpt872__test",
335
+ "benchmark": "spider_realistic",
336
+ "seed": 100,
337
+ "scores": {
338
+ "Exact Match": 0.449,
339
+ "Execution Accuracy": 0.502
340
+ },
341
+ "artifacts": {
342
+ "run_name": "baselines__train_rkl__ckpt872__test",
343
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
344
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
345
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
346
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
347
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
348
+ "ckpt_step": "872"
349
+ }
350
+ },
351
+ {
352
+ "run_name": "baselines__train_rkl__ckpt872__test",
353
+ "benchmark": "spider_syn",
354
+ "seed": 100,
355
+ "scores": {
356
+ "Exact Match": 0.41,
357
+ "Execution Accuracy": 0.479
358
+ },
359
+ "artifacts": {
360
+ "run_name": "baselines__train_rkl__ckpt872__test",
361
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
362
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
363
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
364
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
365
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
366
+ "ckpt_step": "872"
367
+ }
368
+ },
369
+ {
370
+ "run_name": "baselines__train_sfkl__ckpt654__test",
371
+ "benchmark": "spider_data",
372
+ "seed": 100,
373
+ "scores": {
374
+ "Exact Match": 0.51,
375
+ "Execution Accuracy": 0.577
376
+ },
377
+ "artifacts": {
378
+ "run_name": "baselines__train_sfkl__ckpt654__test",
379
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
380
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
381
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
382
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
383
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
384
+ "ckpt_step": "654"
385
+ }
386
+ },
387
+ {
388
+ "run_name": "baselines__train_sfkl__ckpt654__test",
389
+ "benchmark": "spider_dk",
390
+ "seed": 100,
391
+ "scores": {
392
+ "Exact Match": 0.404,
393
+ "Execution Accuracy": 0.452
394
+ },
395
+ "artifacts": {
396
+ "run_name": "baselines__train_sfkl__ckpt654__test",
397
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
398
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
399
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
400
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
401
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
402
+ "ckpt_step": "654"
403
+ }
404
+ },
405
+ {
406
+ "run_name": "baselines__train_sfkl__ckpt654__test",
407
+ "benchmark": "spider_realistic",
408
+ "seed": 100,
409
+ "scores": {
410
+ "Exact Match": 0.441,
411
+ "Execution Accuracy": 0.49
412
+ },
413
+ "artifacts": {
414
+ "run_name": "baselines__train_sfkl__ckpt654__test",
415
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
416
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
417
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
418
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
419
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
420
+ "ckpt_step": "654"
421
+ }
422
+ },
423
+ {
424
+ "run_name": "baselines__train_sfkl__ckpt654__test",
425
+ "benchmark": "spider_syn",
426
+ "seed": 100,
427
+ "scores": {
428
+ "Exact Match": 0.392,
429
+ "Execution Accuracy": 0.436
430
+ },
431
+ "artifacts": {
432
+ "run_name": "baselines__train_sfkl__ckpt654__test",
433
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
434
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
435
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
436
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
437
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
438
+ "ckpt_step": "654"
439
+ }
440
+ },
441
+ {
442
+ "run_name": "baselines__train_srkl__ckpt436__test",
443
+ "benchmark": "spider_data",
444
+ "seed": 100,
445
+ "scores": {
446
+ "Exact Match": 0.487,
447
+ "Execution Accuracy": 0.552
448
+ },
449
+ "artifacts": {
450
+ "run_name": "baselines__train_srkl__ckpt436__test",
451
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
452
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed100/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
453
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
454
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
455
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
456
+ "ckpt_step": "436"
457
+ }
458
+ },
459
+ {
460
+ "run_name": "baselines__train_srkl__ckpt436__test",
461
+ "benchmark": "spider_dk",
462
+ "seed": 100,
463
+ "scores": {
464
+ "Exact Match": 0.342,
465
+ "Execution Accuracy": 0.413
466
+ },
467
+ "artifacts": {
468
+ "run_name": "baselines__train_srkl__ckpt436__test",
469
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
470
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed100/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
471
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed100/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
472
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
473
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
474
+ "ckpt_step": "436"
475
+ }
476
+ },
477
+ {
478
+ "run_name": "baselines__train_srkl__ckpt436__test",
479
+ "benchmark": "spider_realistic",
480
+ "seed": 100,
481
+ "scores": {
482
+ "Exact Match": 0.394,
483
+ "Execution Accuracy": 0.457
484
+ },
485
+ "artifacts": {
486
+ "run_name": "baselines__train_srkl__ckpt436__test",
487
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
488
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed100/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
489
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed100/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
490
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
491
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
492
+ "ckpt_step": "436"
493
+ }
494
+ },
495
+ {
496
+ "run_name": "baselines__train_srkl__ckpt436__test",
497
+ "benchmark": "spider_syn",
498
+ "seed": 100,
499
+ "scores": {
500
+ "Exact Match": 0.348,
501
+ "Execution Accuracy": 0.407
502
+ },
503
+ "artifacts": {
504
+ "run_name": "baselines__train_srkl__ckpt436__test",
505
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
506
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed100/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
507
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed100/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
508
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
509
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
510
+ "ckpt_step": "436"
511
+ }
512
+ }
513
+ ]
514
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed1234/best_by_benchmark.json ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 1234,
3
+ "generated_at": "2026-07-20T12:37:18.614635+00:00",
4
+ "primary_metric": "Execution Accuracy",
5
+ "tie_breaker": "Exact Match",
6
+ "best_by_benchmark": {
7
+ "spider_data": {
8
+ "run_name": "baselines__train_rkl__ckpt872__test",
9
+ "scores": {
10
+ "Exact Match": 0.55,
11
+ "Execution Accuracy": 0.612
12
+ },
13
+ "artifacts": {
14
+ "run_name": "baselines__train_rkl__ckpt872__test",
15
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
16
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
17
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
18
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
19
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
20
+ "ckpt_step": "872"
21
+ }
22
+ },
23
+ "spider_dk": {
24
+ "run_name": "baselines__train_csd__ckpt654__test",
25
+ "scores": {
26
+ "Exact Match": 0.426,
27
+ "Execution Accuracy": 0.484
28
+ },
29
+ "artifacts": {
30
+ "run_name": "baselines__train_csd__ckpt654__test",
31
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
32
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
33
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
34
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
35
+ "train_script": "scripts/llama/baselines/train_csd.sh",
36
+ "ckpt_step": "654"
37
+ }
38
+ },
39
+ "spider_realistic": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "scores": {
42
+ "Exact Match": 0.447,
43
+ "Execution Accuracy": 0.492
44
+ },
45
+ "artifacts": {
46
+ "run_name": "baselines__train_rkl__ckpt872__test",
47
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
48
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
49
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
50
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
51
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
52
+ "ckpt_step": "872"
53
+ }
54
+ },
55
+ "spider_syn": {
56
+ "run_name": "baselines__train_rkl__ckpt872__test",
57
+ "scores": {
58
+ "Exact Match": 0.407,
59
+ "Execution Accuracy": 0.475
60
+ },
61
+ "artifacts": {
62
+ "run_name": "baselines__train_rkl__ckpt872__test",
63
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
64
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
65
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
66
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
67
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
68
+ "ckpt_step": "872"
69
+ }
70
+ }
71
+ }
72
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed1234/best_overall.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 1234,
3
+ "generated_at": "2026-07-20T12:37:18.614635+00:00",
4
+ "aggregation": {
5
+ "benchmarks": [
6
+ "spider_data",
7
+ "spider_syn",
8
+ "spider_realistic",
9
+ "spider_dk"
10
+ ],
11
+ "primary_metric": "mean Execution Accuracy",
12
+ "tie_breaker": "mean Exact Match"
13
+ },
14
+ "best": {
15
+ "run_name": "baselines__train_csd__ckpt654__test",
16
+ "mean_scores": {
17
+ "Exact Match": 0.45075,
18
+ "Execution Accuracy": 0.5065
19
+ },
20
+ "per_benchmark": {
21
+ "spider_data": {
22
+ "Exact Match": 0.536,
23
+ "Execution Accuracy": 0.599
24
+ },
25
+ "spider_syn": {
26
+ "Exact Match": 0.408,
27
+ "Execution Accuracy": 0.465
28
+ },
29
+ "spider_realistic": {
30
+ "Exact Match": 0.433,
31
+ "Execution Accuracy": 0.478
32
+ },
33
+ "spider_dk": {
34
+ "Exact Match": 0.426,
35
+ "Execution Accuracy": 0.484
36
+ }
37
+ },
38
+ "artifacts": {
39
+ "spider_data": {
40
+ "run_name": "baselines__train_csd__ckpt654__test",
41
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
42
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
43
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
44
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
45
+ "train_script": "scripts/llama/baselines/train_csd.sh",
46
+ "ckpt_step": "654"
47
+ },
48
+ "spider_syn": {
49
+ "run_name": "baselines__train_csd__ckpt654__test",
50
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
51
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
52
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
53
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
54
+ "train_script": "scripts/llama/baselines/train_csd.sh",
55
+ "ckpt_step": "654"
56
+ },
57
+ "spider_realistic": {
58
+ "run_name": "baselines__train_csd__ckpt654__test",
59
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
60
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
61
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
62
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
63
+ "train_script": "scripts/llama/baselines/train_csd.sh",
64
+ "ckpt_step": "654"
65
+ },
66
+ "spider_dk": {
67
+ "run_name": "baselines__train_csd__ckpt654__test",
68
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
69
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
70
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
71
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
72
+ "train_script": "scripts/llama/baselines/train_csd.sh",
73
+ "ckpt_step": "654"
74
+ }
75
+ }
76
+ }
77
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed1234/eval_results.json ADDED
@@ -0,0 +1,514 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 1234,
3
+ "generated_at": "2026-07-20T12:37:18.614635+00:00",
4
+ "selection": {
5
+ "primary_metric": "Execution Accuracy",
6
+ "tie_breaker": "Exact Match"
7
+ },
8
+ "results": [
9
+ {
10
+ "run_name": "baselines__train_amid__ckpt872__test",
11
+ "benchmark": "spider_data",
12
+ "seed": 1234,
13
+ "scores": {
14
+ "Exact Match": 0.509,
15
+ "Execution Accuracy": 0.573
16
+ },
17
+ "artifacts": {
18
+ "run_name": "baselines__train_amid__ckpt872__test",
19
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
20
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
21
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
22
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
23
+ "train_script": "scripts/llama/baselines/train_amid.sh",
24
+ "ckpt_step": "872"
25
+ }
26
+ },
27
+ {
28
+ "run_name": "baselines__train_amid__ckpt872__test",
29
+ "benchmark": "spider_dk",
30
+ "seed": 1234,
31
+ "scores": {
32
+ "Exact Match": 0.394,
33
+ "Execution Accuracy": 0.475
34
+ },
35
+ "artifacts": {
36
+ "run_name": "baselines__train_amid__ckpt872__test",
37
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
38
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
39
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
40
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
41
+ "train_script": "scripts/llama/baselines/train_amid.sh",
42
+ "ckpt_step": "872"
43
+ }
44
+ },
45
+ {
46
+ "run_name": "baselines__train_amid__ckpt872__test",
47
+ "benchmark": "spider_realistic",
48
+ "seed": 1234,
49
+ "scores": {
50
+ "Exact Match": 0.417,
51
+ "Execution Accuracy": 0.48
52
+ },
53
+ "artifacts": {
54
+ "run_name": "baselines__train_amid__ckpt872__test",
55
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
56
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
57
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
58
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
59
+ "train_script": "scripts/llama/baselines/train_amid.sh",
60
+ "ckpt_step": "872"
61
+ }
62
+ },
63
+ {
64
+ "run_name": "baselines__train_amid__ckpt872__test",
65
+ "benchmark": "spider_syn",
66
+ "seed": 1234,
67
+ "scores": {
68
+ "Exact Match": 0.356,
69
+ "Execution Accuracy": 0.421
70
+ },
71
+ "artifacts": {
72
+ "run_name": "baselines__train_amid__ckpt872__test",
73
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
74
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
75
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
76
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
77
+ "train_script": "scripts/llama/baselines/train_amid.sh",
78
+ "ckpt_step": "872"
79
+ }
80
+ },
81
+ {
82
+ "run_name": "baselines__train_csd__ckpt654__test",
83
+ "benchmark": "spider_data",
84
+ "seed": 1234,
85
+ "scores": {
86
+ "Exact Match": 0.536,
87
+ "Execution Accuracy": 0.599
88
+ },
89
+ "artifacts": {
90
+ "run_name": "baselines__train_csd__ckpt654__test",
91
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
92
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
93
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
94
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
95
+ "train_script": "scripts/llama/baselines/train_csd.sh",
96
+ "ckpt_step": "654"
97
+ }
98
+ },
99
+ {
100
+ "run_name": "baselines__train_csd__ckpt654__test",
101
+ "benchmark": "spider_dk",
102
+ "seed": 1234,
103
+ "scores": {
104
+ "Exact Match": 0.426,
105
+ "Execution Accuracy": 0.484
106
+ },
107
+ "artifacts": {
108
+ "run_name": "baselines__train_csd__ckpt654__test",
109
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
110
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
111
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
112
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
113
+ "train_script": "scripts/llama/baselines/train_csd.sh",
114
+ "ckpt_step": "654"
115
+ }
116
+ },
117
+ {
118
+ "run_name": "baselines__train_csd__ckpt654__test",
119
+ "benchmark": "spider_realistic",
120
+ "seed": 1234,
121
+ "scores": {
122
+ "Exact Match": 0.433,
123
+ "Execution Accuracy": 0.478
124
+ },
125
+ "artifacts": {
126
+ "run_name": "baselines__train_csd__ckpt654__test",
127
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
128
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
129
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
130
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
131
+ "train_script": "scripts/llama/baselines/train_csd.sh",
132
+ "ckpt_step": "654"
133
+ }
134
+ },
135
+ {
136
+ "run_name": "baselines__train_csd__ckpt654__test",
137
+ "benchmark": "spider_syn",
138
+ "seed": 1234,
139
+ "scores": {
140
+ "Exact Match": 0.408,
141
+ "Execution Accuracy": 0.465
142
+ },
143
+ "artifacts": {
144
+ "run_name": "baselines__train_csd__ckpt654__test",
145
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
146
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
147
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
148
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
149
+ "train_script": "scripts/llama/baselines/train_csd.sh",
150
+ "ckpt_step": "654"
151
+ }
152
+ },
153
+ {
154
+ "run_name": "baselines__train_distillm__ckpt654__test",
155
+ "benchmark": "spider_data",
156
+ "seed": 1234,
157
+ "scores": {
158
+ "Exact Match": 0.504,
159
+ "Execution Accuracy": 0.574
160
+ },
161
+ "artifacts": {
162
+ "run_name": "baselines__train_distillm__ckpt654__test",
163
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
164
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
165
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
166
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
167
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
168
+ "ckpt_step": "654"
169
+ }
170
+ },
171
+ {
172
+ "run_name": "baselines__train_distillm__ckpt654__test",
173
+ "benchmark": "spider_dk",
174
+ "seed": 1234,
175
+ "scores": {
176
+ "Exact Match": 0.394,
177
+ "Execution Accuracy": 0.46
178
+ },
179
+ "artifacts": {
180
+ "run_name": "baselines__train_distillm__ckpt654__test",
181
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
182
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
183
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
184
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
185
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
186
+ "ckpt_step": "654"
187
+ }
188
+ },
189
+ {
190
+ "run_name": "baselines__train_distillm__ckpt654__test",
191
+ "benchmark": "spider_realistic",
192
+ "seed": 1234,
193
+ "scores": {
194
+ "Exact Match": 0.421,
195
+ "Execution Accuracy": 0.47
196
+ },
197
+ "artifacts": {
198
+ "run_name": "baselines__train_distillm__ckpt654__test",
199
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
200
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
201
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
202
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
203
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
204
+ "ckpt_step": "654"
205
+ }
206
+ },
207
+ {
208
+ "run_name": "baselines__train_distillm__ckpt654__test",
209
+ "benchmark": "spider_syn",
210
+ "seed": 1234,
211
+ "scores": {
212
+ "Exact Match": 0.381,
213
+ "Execution Accuracy": 0.446
214
+ },
215
+ "artifacts": {
216
+ "run_name": "baselines__train_distillm__ckpt654__test",
217
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
218
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
219
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
220
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
221
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
222
+ "ckpt_step": "654"
223
+ }
224
+ },
225
+ {
226
+ "run_name": "baselines__train_fkl__ckpt872__test",
227
+ "benchmark": "spider_data",
228
+ "seed": 1234,
229
+ "scores": {
230
+ "Exact Match": 0.515,
231
+ "Execution Accuracy": 0.578
232
+ },
233
+ "artifacts": {
234
+ "run_name": "baselines__train_fkl__ckpt872__test",
235
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
236
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
237
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
238
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
239
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
240
+ "ckpt_step": "872"
241
+ }
242
+ },
243
+ {
244
+ "run_name": "baselines__train_fkl__ckpt872__test",
245
+ "benchmark": "spider_dk",
246
+ "seed": 1234,
247
+ "scores": {
248
+ "Exact Match": 0.374,
249
+ "Execution Accuracy": 0.436
250
+ },
251
+ "artifacts": {
252
+ "run_name": "baselines__train_fkl__ckpt872__test",
253
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
254
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
255
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
256
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
257
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
258
+ "ckpt_step": "872"
259
+ }
260
+ },
261
+ {
262
+ "run_name": "baselines__train_fkl__ckpt872__test",
263
+ "benchmark": "spider_realistic",
264
+ "seed": 1234,
265
+ "scores": {
266
+ "Exact Match": 0.411,
267
+ "Execution Accuracy": 0.453
268
+ },
269
+ "artifacts": {
270
+ "run_name": "baselines__train_fkl__ckpt872__test",
271
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
272
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
273
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
274
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
275
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
276
+ "ckpt_step": "872"
277
+ }
278
+ },
279
+ {
280
+ "run_name": "baselines__train_fkl__ckpt872__test",
281
+ "benchmark": "spider_syn",
282
+ "seed": 1234,
283
+ "scores": {
284
+ "Exact Match": 0.359,
285
+ "Execution Accuracy": 0.424
286
+ },
287
+ "artifacts": {
288
+ "run_name": "baselines__train_fkl__ckpt872__test",
289
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
290
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
291
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
292
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
293
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
294
+ "ckpt_step": "872"
295
+ }
296
+ },
297
+ {
298
+ "run_name": "baselines__train_rkl__ckpt872__test",
299
+ "benchmark": "spider_data",
300
+ "seed": 1234,
301
+ "scores": {
302
+ "Exact Match": 0.55,
303
+ "Execution Accuracy": 0.612
304
+ },
305
+ "artifacts": {
306
+ "run_name": "baselines__train_rkl__ckpt872__test",
307
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
308
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
309
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
310
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
311
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
312
+ "ckpt_step": "872"
313
+ }
314
+ },
315
+ {
316
+ "run_name": "baselines__train_rkl__ckpt872__test",
317
+ "benchmark": "spider_dk",
318
+ "seed": 1234,
319
+ "scores": {
320
+ "Exact Match": 0.394,
321
+ "Execution Accuracy": 0.441
322
+ },
323
+ "artifacts": {
324
+ "run_name": "baselines__train_rkl__ckpt872__test",
325
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
326
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
327
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
328
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
329
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
330
+ "ckpt_step": "872"
331
+ }
332
+ },
333
+ {
334
+ "run_name": "baselines__train_rkl__ckpt872__test",
335
+ "benchmark": "spider_realistic",
336
+ "seed": 1234,
337
+ "scores": {
338
+ "Exact Match": 0.447,
339
+ "Execution Accuracy": 0.492
340
+ },
341
+ "artifacts": {
342
+ "run_name": "baselines__train_rkl__ckpt872__test",
343
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
344
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
345
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
346
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
347
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
348
+ "ckpt_step": "872"
349
+ }
350
+ },
351
+ {
352
+ "run_name": "baselines__train_rkl__ckpt872__test",
353
+ "benchmark": "spider_syn",
354
+ "seed": 1234,
355
+ "scores": {
356
+ "Exact Match": 0.407,
357
+ "Execution Accuracy": 0.475
358
+ },
359
+ "artifacts": {
360
+ "run_name": "baselines__train_rkl__ckpt872__test",
361
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
362
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
363
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
364
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
365
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
366
+ "ckpt_step": "872"
367
+ }
368
+ },
369
+ {
370
+ "run_name": "baselines__train_sfkl__ckpt654__test",
371
+ "benchmark": "spider_data",
372
+ "seed": 1234,
373
+ "scores": {
374
+ "Exact Match": 0.507,
375
+ "Execution Accuracy": 0.578
376
+ },
377
+ "artifacts": {
378
+ "run_name": "baselines__train_sfkl__ckpt654__test",
379
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
380
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
381
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
382
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
383
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
384
+ "ckpt_step": "654"
385
+ }
386
+ },
387
+ {
388
+ "run_name": "baselines__train_sfkl__ckpt654__test",
389
+ "benchmark": "spider_dk",
390
+ "seed": 1234,
391
+ "scores": {
392
+ "Exact Match": 0.391,
393
+ "Execution Accuracy": 0.454
394
+ },
395
+ "artifacts": {
396
+ "run_name": "baselines__train_sfkl__ckpt654__test",
397
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
398
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
399
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
400
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
401
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
402
+ "ckpt_step": "654"
403
+ }
404
+ },
405
+ {
406
+ "run_name": "baselines__train_sfkl__ckpt654__test",
407
+ "benchmark": "spider_realistic",
408
+ "seed": 1234,
409
+ "scores": {
410
+ "Exact Match": 0.425,
411
+ "Execution Accuracy": 0.463
412
+ },
413
+ "artifacts": {
414
+ "run_name": "baselines__train_sfkl__ckpt654__test",
415
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
416
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
417
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
418
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
419
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
420
+ "ckpt_step": "654"
421
+ }
422
+ },
423
+ {
424
+ "run_name": "baselines__train_sfkl__ckpt654__test",
425
+ "benchmark": "spider_syn",
426
+ "seed": 1234,
427
+ "scores": {
428
+ "Exact Match": 0.378,
429
+ "Execution Accuracy": 0.416
430
+ },
431
+ "artifacts": {
432
+ "run_name": "baselines__train_sfkl__ckpt654__test",
433
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
434
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
435
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
436
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
437
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
438
+ "ckpt_step": "654"
439
+ }
440
+ },
441
+ {
442
+ "run_name": "baselines__train_srkl__ckpt436__test",
443
+ "benchmark": "spider_data",
444
+ "seed": 1234,
445
+ "scores": {
446
+ "Exact Match": 0.47,
447
+ "Execution Accuracy": 0.543
448
+ },
449
+ "artifacts": {
450
+ "run_name": "baselines__train_srkl__ckpt436__test",
451
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
452
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
453
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
454
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
455
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
456
+ "ckpt_step": "436"
457
+ }
458
+ },
459
+ {
460
+ "run_name": "baselines__train_srkl__ckpt436__test",
461
+ "benchmark": "spider_dk",
462
+ "seed": 1234,
463
+ "scores": {
464
+ "Exact Match": 0.348,
465
+ "Execution Accuracy": 0.419
466
+ },
467
+ "artifacts": {
468
+ "run_name": "baselines__train_srkl__ckpt436__test",
469
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
470
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
471
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed1234/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
472
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
473
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
474
+ "ckpt_step": "436"
475
+ }
476
+ },
477
+ {
478
+ "run_name": "baselines__train_srkl__ckpt436__test",
479
+ "benchmark": "spider_realistic",
480
+ "seed": 1234,
481
+ "scores": {
482
+ "Exact Match": 0.402,
483
+ "Execution Accuracy": 0.453
484
+ },
485
+ "artifacts": {
486
+ "run_name": "baselines__train_srkl__ckpt436__test",
487
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
488
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
489
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed1234/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
490
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
491
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
492
+ "ckpt_step": "436"
493
+ }
494
+ },
495
+ {
496
+ "run_name": "baselines__train_srkl__ckpt436__test",
497
+ "benchmark": "spider_syn",
498
+ "seed": 1234,
499
+ "scores": {
500
+ "Exact Match": 0.353,
501
+ "Execution Accuracy": 0.406
502
+ },
503
+ "artifacts": {
504
+ "run_name": "baselines__train_srkl__ckpt436__test",
505
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
506
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed1234/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
507
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed1234/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
508
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
509
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
510
+ "ckpt_step": "436"
511
+ }
512
+ }
513
+ ]
514
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed42/best_by_benchmark.json ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 42,
3
+ "generated_at": "2026-07-20T12:37:18.611321+00:00",
4
+ "primary_metric": "Execution Accuracy",
5
+ "tie_breaker": "Exact Match",
6
+ "best_by_benchmark": {
7
+ "spider_data": {
8
+ "run_name": "baselines__train_rkl__ckpt872__test",
9
+ "scores": {
10
+ "Exact Match": 0.549,
11
+ "Execution Accuracy": 0.611
12
+ },
13
+ "artifacts": {
14
+ "run_name": "baselines__train_rkl__ckpt872__test",
15
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
16
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
17
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
18
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
19
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
20
+ "ckpt_step": "872"
21
+ }
22
+ },
23
+ "spider_dk": {
24
+ "run_name": "baselines__train_csd__ckpt654__test",
25
+ "scores": {
26
+ "Exact Match": 0.413,
27
+ "Execution Accuracy": 0.477
28
+ },
29
+ "artifacts": {
30
+ "run_name": "baselines__train_csd__ckpt654__test",
31
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
32
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
33
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
34
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
35
+ "train_script": "scripts/llama/baselines/train_csd.sh",
36
+ "ckpt_step": "654"
37
+ }
38
+ },
39
+ "spider_realistic": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "scores": {
42
+ "Exact Match": 0.461,
43
+ "Execution Accuracy": 0.504
44
+ },
45
+ "artifacts": {
46
+ "run_name": "baselines__train_rkl__ckpt872__test",
47
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
48
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
49
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
50
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
51
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
52
+ "ckpt_step": "872"
53
+ }
54
+ },
55
+ "spider_syn": {
56
+ "run_name": "baselines__train_rkl__ckpt872__test",
57
+ "scores": {
58
+ "Exact Match": 0.405,
59
+ "Execution Accuracy": 0.476
60
+ },
61
+ "artifacts": {
62
+ "run_name": "baselines__train_rkl__ckpt872__test",
63
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
64
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
65
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
66
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
67
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
68
+ "ckpt_step": "872"
69
+ }
70
+ }
71
+ }
72
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed42/best_overall.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 42,
3
+ "generated_at": "2026-07-20T12:37:18.611321+00:00",
4
+ "aggregation": {
5
+ "benchmarks": [
6
+ "spider_data",
7
+ "spider_syn",
8
+ "spider_realistic",
9
+ "spider_dk"
10
+ ],
11
+ "primary_metric": "mean Execution Accuracy",
12
+ "tie_breaker": "mean Exact Match"
13
+ },
14
+ "best": {
15
+ "run_name": "baselines__train_rkl__ckpt872__test",
16
+ "mean_scores": {
17
+ "Exact Match": 0.4505,
18
+ "Execution Accuracy": 0.508
19
+ },
20
+ "per_benchmark": {
21
+ "spider_data": {
22
+ "Exact Match": 0.549,
23
+ "Execution Accuracy": 0.611
24
+ },
25
+ "spider_syn": {
26
+ "Exact Match": 0.405,
27
+ "Execution Accuracy": 0.476
28
+ },
29
+ "spider_realistic": {
30
+ "Exact Match": 0.461,
31
+ "Execution Accuracy": 0.504
32
+ },
33
+ "spider_dk": {
34
+ "Exact Match": 0.387,
35
+ "Execution Accuracy": 0.441
36
+ }
37
+ },
38
+ "artifacts": {
39
+ "spider_data": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
42
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
43
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
44
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
45
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
46
+ "ckpt_step": "872"
47
+ },
48
+ "spider_syn": {
49
+ "run_name": "baselines__train_rkl__ckpt872__test",
50
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
51
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
52
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
53
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
54
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
55
+ "ckpt_step": "872"
56
+ },
57
+ "spider_realistic": {
58
+ "run_name": "baselines__train_rkl__ckpt872__test",
59
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
60
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
61
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
62
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
63
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
64
+ "ckpt_step": "872"
65
+ },
66
+ "spider_dk": {
67
+ "run_name": "baselines__train_rkl__ckpt872__test",
68
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
69
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
70
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
71
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
72
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
73
+ "ckpt_step": "872"
74
+ }
75
+ }
76
+ }
77
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed42/eval_results.json ADDED
@@ -0,0 +1,514 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 42,
3
+ "generated_at": "2026-07-20T12:37:18.611321+00:00",
4
+ "selection": {
5
+ "primary_metric": "Execution Accuracy",
6
+ "tie_breaker": "Exact Match"
7
+ },
8
+ "results": [
9
+ {
10
+ "run_name": "baselines__train_amid__ckpt872__test",
11
+ "benchmark": "spider_data",
12
+ "seed": 42,
13
+ "scores": {
14
+ "Exact Match": 0.51,
15
+ "Execution Accuracy": 0.58
16
+ },
17
+ "artifacts": {
18
+ "run_name": "baselines__train_amid__ckpt872__test",
19
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
20
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
21
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
22
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
23
+ "train_script": "scripts/llama/baselines/train_amid.sh",
24
+ "ckpt_step": "872"
25
+ }
26
+ },
27
+ {
28
+ "run_name": "baselines__train_amid__ckpt872__test",
29
+ "benchmark": "spider_dk",
30
+ "seed": 42,
31
+ "scores": {
32
+ "Exact Match": 0.379,
33
+ "Execution Accuracy": 0.45
34
+ },
35
+ "artifacts": {
36
+ "run_name": "baselines__train_amid__ckpt872__test",
37
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
38
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
39
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
40
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
41
+ "train_script": "scripts/llama/baselines/train_amid.sh",
42
+ "ckpt_step": "872"
43
+ }
44
+ },
45
+ {
46
+ "run_name": "baselines__train_amid__ckpt872__test",
47
+ "benchmark": "spider_realistic",
48
+ "seed": 42,
49
+ "scores": {
50
+ "Exact Match": 0.421,
51
+ "Execution Accuracy": 0.47
52
+ },
53
+ "artifacts": {
54
+ "run_name": "baselines__train_amid__ckpt872__test",
55
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
56
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
57
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
58
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
59
+ "train_script": "scripts/llama/baselines/train_amid.sh",
60
+ "ckpt_step": "872"
61
+ }
62
+ },
63
+ {
64
+ "run_name": "baselines__train_amid__ckpt872__test",
65
+ "benchmark": "spider_syn",
66
+ "seed": 42,
67
+ "scores": {
68
+ "Exact Match": 0.356,
69
+ "Execution Accuracy": 0.428
70
+ },
71
+ "artifacts": {
72
+ "run_name": "baselines__train_amid__ckpt872__test",
73
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
74
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
75
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
76
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
77
+ "train_script": "scripts/llama/baselines/train_amid.sh",
78
+ "ckpt_step": "872"
79
+ }
80
+ },
81
+ {
82
+ "run_name": "baselines__train_csd__ckpt654__test",
83
+ "benchmark": "spider_data",
84
+ "seed": 42,
85
+ "scores": {
86
+ "Exact Match": 0.534,
87
+ "Execution Accuracy": 0.589
88
+ },
89
+ "artifacts": {
90
+ "run_name": "baselines__train_csd__ckpt654__test",
91
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
92
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
93
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
94
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
95
+ "train_script": "scripts/llama/baselines/train_csd.sh",
96
+ "ckpt_step": "654"
97
+ }
98
+ },
99
+ {
100
+ "run_name": "baselines__train_csd__ckpt654__test",
101
+ "benchmark": "spider_dk",
102
+ "seed": 42,
103
+ "scores": {
104
+ "Exact Match": 0.413,
105
+ "Execution Accuracy": 0.477
106
+ },
107
+ "artifacts": {
108
+ "run_name": "baselines__train_csd__ckpt654__test",
109
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
110
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
111
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
112
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
113
+ "train_script": "scripts/llama/baselines/train_csd.sh",
114
+ "ckpt_step": "654"
115
+ }
116
+ },
117
+ {
118
+ "run_name": "baselines__train_csd__ckpt654__test",
119
+ "benchmark": "spider_realistic",
120
+ "seed": 42,
121
+ "scores": {
122
+ "Exact Match": 0.445,
123
+ "Execution Accuracy": 0.486
124
+ },
125
+ "artifacts": {
126
+ "run_name": "baselines__train_csd__ckpt654__test",
127
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
128
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
129
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
130
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
131
+ "train_script": "scripts/llama/baselines/train_csd.sh",
132
+ "ckpt_step": "654"
133
+ }
134
+ },
135
+ {
136
+ "run_name": "baselines__train_csd__ckpt654__test",
137
+ "benchmark": "spider_syn",
138
+ "seed": 42,
139
+ "scores": {
140
+ "Exact Match": 0.407,
141
+ "Execution Accuracy": 0.463
142
+ },
143
+ "artifacts": {
144
+ "run_name": "baselines__train_csd__ckpt654__test",
145
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
146
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
147
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
148
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
149
+ "train_script": "scripts/llama/baselines/train_csd.sh",
150
+ "ckpt_step": "654"
151
+ }
152
+ },
153
+ {
154
+ "run_name": "baselines__train_distillm__ckpt654__test",
155
+ "benchmark": "spider_data",
156
+ "seed": 42,
157
+ "scores": {
158
+ "Exact Match": 0.503,
159
+ "Execution Accuracy": 0.583
160
+ },
161
+ "artifacts": {
162
+ "run_name": "baselines__train_distillm__ckpt654__test",
163
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
164
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
165
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
166
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
167
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
168
+ "ckpt_step": "654"
169
+ }
170
+ },
171
+ {
172
+ "run_name": "baselines__train_distillm__ckpt654__test",
173
+ "benchmark": "spider_dk",
174
+ "seed": 42,
175
+ "scores": {
176
+ "Exact Match": 0.387,
177
+ "Execution Accuracy": 0.447
178
+ },
179
+ "artifacts": {
180
+ "run_name": "baselines__train_distillm__ckpt654__test",
181
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
182
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
183
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
184
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
185
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
186
+ "ckpt_step": "654"
187
+ }
188
+ },
189
+ {
190
+ "run_name": "baselines__train_distillm__ckpt654__test",
191
+ "benchmark": "spider_realistic",
192
+ "seed": 42,
193
+ "scores": {
194
+ "Exact Match": 0.413,
195
+ "Execution Accuracy": 0.472
196
+ },
197
+ "artifacts": {
198
+ "run_name": "baselines__train_distillm__ckpt654__test",
199
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
200
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
201
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
202
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
203
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
204
+ "ckpt_step": "654"
205
+ }
206
+ },
207
+ {
208
+ "run_name": "baselines__train_distillm__ckpt654__test",
209
+ "benchmark": "spider_syn",
210
+ "seed": 42,
211
+ "scores": {
212
+ "Exact Match": 0.382,
213
+ "Execution Accuracy": 0.435
214
+ },
215
+ "artifacts": {
216
+ "run_name": "baselines__train_distillm__ckpt654__test",
217
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
218
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
219
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
220
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
221
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
222
+ "ckpt_step": "654"
223
+ }
224
+ },
225
+ {
226
+ "run_name": "baselines__train_fkl__ckpt872__test",
227
+ "benchmark": "spider_data",
228
+ "seed": 42,
229
+ "scores": {
230
+ "Exact Match": 0.511,
231
+ "Execution Accuracy": 0.585
232
+ },
233
+ "artifacts": {
234
+ "run_name": "baselines__train_fkl__ckpt872__test",
235
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
236
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
237
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
238
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
239
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
240
+ "ckpt_step": "872"
241
+ }
242
+ },
243
+ {
244
+ "run_name": "baselines__train_fkl__ckpt872__test",
245
+ "benchmark": "spider_dk",
246
+ "seed": 42,
247
+ "scores": {
248
+ "Exact Match": 0.374,
249
+ "Execution Accuracy": 0.458
250
+ },
251
+ "artifacts": {
252
+ "run_name": "baselines__train_fkl__ckpt872__test",
253
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
254
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
255
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
256
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
257
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
258
+ "ckpt_step": "872"
259
+ }
260
+ },
261
+ {
262
+ "run_name": "baselines__train_fkl__ckpt872__test",
263
+ "benchmark": "spider_realistic",
264
+ "seed": 42,
265
+ "scores": {
266
+ "Exact Match": 0.411,
267
+ "Execution Accuracy": 0.461
268
+ },
269
+ "artifacts": {
270
+ "run_name": "baselines__train_fkl__ckpt872__test",
271
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
272
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
273
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
274
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
275
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
276
+ "ckpt_step": "872"
277
+ }
278
+ },
279
+ {
280
+ "run_name": "baselines__train_fkl__ckpt872__test",
281
+ "benchmark": "spider_syn",
282
+ "seed": 42,
283
+ "scores": {
284
+ "Exact Match": 0.367,
285
+ "Execution Accuracy": 0.436
286
+ },
287
+ "artifacts": {
288
+ "run_name": "baselines__train_fkl__ckpt872__test",
289
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
290
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
291
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
292
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
293
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
294
+ "ckpt_step": "872"
295
+ }
296
+ },
297
+ {
298
+ "run_name": "baselines__train_rkl__ckpt872__test",
299
+ "benchmark": "spider_data",
300
+ "seed": 42,
301
+ "scores": {
302
+ "Exact Match": 0.549,
303
+ "Execution Accuracy": 0.611
304
+ },
305
+ "artifacts": {
306
+ "run_name": "baselines__train_rkl__ckpt872__test",
307
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
308
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
309
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
310
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
311
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
312
+ "ckpt_step": "872"
313
+ }
314
+ },
315
+ {
316
+ "run_name": "baselines__train_rkl__ckpt872__test",
317
+ "benchmark": "spider_dk",
318
+ "seed": 42,
319
+ "scores": {
320
+ "Exact Match": 0.387,
321
+ "Execution Accuracy": 0.441
322
+ },
323
+ "artifacts": {
324
+ "run_name": "baselines__train_rkl__ckpt872__test",
325
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
326
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
327
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
328
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
329
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
330
+ "ckpt_step": "872"
331
+ }
332
+ },
333
+ {
334
+ "run_name": "baselines__train_rkl__ckpt872__test",
335
+ "benchmark": "spider_realistic",
336
+ "seed": 42,
337
+ "scores": {
338
+ "Exact Match": 0.461,
339
+ "Execution Accuracy": 0.504
340
+ },
341
+ "artifacts": {
342
+ "run_name": "baselines__train_rkl__ckpt872__test",
343
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
344
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
345
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
346
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
347
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
348
+ "ckpt_step": "872"
349
+ }
350
+ },
351
+ {
352
+ "run_name": "baselines__train_rkl__ckpt872__test",
353
+ "benchmark": "spider_syn",
354
+ "seed": 42,
355
+ "scores": {
356
+ "Exact Match": 0.405,
357
+ "Execution Accuracy": 0.476
358
+ },
359
+ "artifacts": {
360
+ "run_name": "baselines__train_rkl__ckpt872__test",
361
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
362
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
363
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
364
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
365
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
366
+ "ckpt_step": "872"
367
+ }
368
+ },
369
+ {
370
+ "run_name": "baselines__train_sfkl__ckpt654__test",
371
+ "benchmark": "spider_data",
372
+ "seed": 42,
373
+ "scores": {
374
+ "Exact Match": 0.513,
375
+ "Execution Accuracy": 0.578
376
+ },
377
+ "artifacts": {
378
+ "run_name": "baselines__train_sfkl__ckpt654__test",
379
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
380
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
381
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
382
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
383
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
384
+ "ckpt_step": "654"
385
+ }
386
+ },
387
+ {
388
+ "run_name": "baselines__train_sfkl__ckpt654__test",
389
+ "benchmark": "spider_dk",
390
+ "seed": 42,
391
+ "scores": {
392
+ "Exact Match": 0.402,
393
+ "Execution Accuracy": 0.467
394
+ },
395
+ "artifacts": {
396
+ "run_name": "baselines__train_sfkl__ckpt654__test",
397
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
398
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
399
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
400
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
401
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
402
+ "ckpt_step": "654"
403
+ }
404
+ },
405
+ {
406
+ "run_name": "baselines__train_sfkl__ckpt654__test",
407
+ "benchmark": "spider_realistic",
408
+ "seed": 42,
409
+ "scores": {
410
+ "Exact Match": 0.461,
411
+ "Execution Accuracy": 0.492
412
+ },
413
+ "artifacts": {
414
+ "run_name": "baselines__train_sfkl__ckpt654__test",
415
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
416
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
417
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
418
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
419
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
420
+ "ckpt_step": "654"
421
+ }
422
+ },
423
+ {
424
+ "run_name": "baselines__train_sfkl__ckpt654__test",
425
+ "benchmark": "spider_syn",
426
+ "seed": 42,
427
+ "scores": {
428
+ "Exact Match": 0.388,
429
+ "Execution Accuracy": 0.424
430
+ },
431
+ "artifacts": {
432
+ "run_name": "baselines__train_sfkl__ckpt654__test",
433
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
434
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
435
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
436
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
437
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
438
+ "ckpt_step": "654"
439
+ }
440
+ },
441
+ {
442
+ "run_name": "baselines__train_srkl__ckpt436__test",
443
+ "benchmark": "spider_data",
444
+ "seed": 42,
445
+ "scores": {
446
+ "Exact Match": 0.468,
447
+ "Execution Accuracy": 0.537
448
+ },
449
+ "artifacts": {
450
+ "run_name": "baselines__train_srkl__ckpt436__test",
451
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
452
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed42/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
453
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
454
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
455
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
456
+ "ckpt_step": "436"
457
+ }
458
+ },
459
+ {
460
+ "run_name": "baselines__train_srkl__ckpt436__test",
461
+ "benchmark": "spider_dk",
462
+ "seed": 42,
463
+ "scores": {
464
+ "Exact Match": 0.336,
465
+ "Execution Accuracy": 0.415
466
+ },
467
+ "artifacts": {
468
+ "run_name": "baselines__train_srkl__ckpt436__test",
469
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
470
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed42/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
471
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed42/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
472
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
473
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
474
+ "ckpt_step": "436"
475
+ }
476
+ },
477
+ {
478
+ "run_name": "baselines__train_srkl__ckpt436__test",
479
+ "benchmark": "spider_realistic",
480
+ "seed": 42,
481
+ "scores": {
482
+ "Exact Match": 0.394,
483
+ "Execution Accuracy": 0.449
484
+ },
485
+ "artifacts": {
486
+ "run_name": "baselines__train_srkl__ckpt436__test",
487
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
488
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed42/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
489
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed42/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
490
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
491
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
492
+ "ckpt_step": "436"
493
+ }
494
+ },
495
+ {
496
+ "run_name": "baselines__train_srkl__ckpt436__test",
497
+ "benchmark": "spider_syn",
498
+ "seed": 42,
499
+ "scores": {
500
+ "Exact Match": 0.357,
501
+ "Execution Accuracy": 0.409
502
+ },
503
+ "artifacts": {
504
+ "run_name": "baselines__train_srkl__ckpt436__test",
505
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
506
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed42/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
507
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed42/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
508
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
509
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
510
+ "ckpt_step": "436"
511
+ }
512
+ }
513
+ ]
514
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed50/best_by_benchmark.json ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 50,
3
+ "generated_at": "2026-07-20T12:37:18.612447+00:00",
4
+ "primary_metric": "Execution Accuracy",
5
+ "tie_breaker": "Exact Match",
6
+ "best_by_benchmark": {
7
+ "spider_data": {
8
+ "run_name": "baselines__train_rkl__ckpt872__test",
9
+ "scores": {
10
+ "Exact Match": 0.547,
11
+ "Execution Accuracy": 0.61
12
+ },
13
+ "artifacts": {
14
+ "run_name": "baselines__train_rkl__ckpt872__test",
15
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
16
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
17
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
18
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
19
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
20
+ "ckpt_step": "872"
21
+ }
22
+ },
23
+ "spider_dk": {
24
+ "run_name": "baselines__train_amid__ckpt872__test",
25
+ "scores": {
26
+ "Exact Match": 0.391,
27
+ "Execution Accuracy": 0.471
28
+ },
29
+ "artifacts": {
30
+ "run_name": "baselines__train_amid__ckpt872__test",
31
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
32
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
33
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
34
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
35
+ "train_script": "scripts/llama/baselines/train_amid.sh",
36
+ "ckpt_step": "872"
37
+ }
38
+ },
39
+ "spider_realistic": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "scores": {
42
+ "Exact Match": 0.457,
43
+ "Execution Accuracy": 0.502
44
+ },
45
+ "artifacts": {
46
+ "run_name": "baselines__train_rkl__ckpt872__test",
47
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
48
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
49
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
50
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
51
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
52
+ "ckpt_step": "872"
53
+ }
54
+ },
55
+ "spider_syn": {
56
+ "run_name": "baselines__train_csd__ckpt654__test",
57
+ "scores": {
58
+ "Exact Match": 0.414,
59
+ "Execution Accuracy": 0.471
60
+ },
61
+ "artifacts": {
62
+ "run_name": "baselines__train_csd__ckpt654__test",
63
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
64
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
65
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
66
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
67
+ "train_script": "scripts/llama/baselines/train_csd.sh",
68
+ "ckpt_step": "654"
69
+ }
70
+ }
71
+ }
72
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed50/best_overall.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 50,
3
+ "generated_at": "2026-07-20T12:37:18.612447+00:00",
4
+ "aggregation": {
5
+ "benchmarks": [
6
+ "spider_data",
7
+ "spider_syn",
8
+ "spider_realistic",
9
+ "spider_dk"
10
+ ],
11
+ "primary_metric": "mean Execution Accuracy",
12
+ "tie_breaker": "mean Exact Match"
13
+ },
14
+ "best": {
15
+ "run_name": "baselines__train_rkl__ckpt872__test",
16
+ "mean_scores": {
17
+ "Exact Match": 0.4545,
18
+ "Execution Accuracy": 0.50775
19
+ },
20
+ "per_benchmark": {
21
+ "spider_data": {
22
+ "Exact Match": 0.547,
23
+ "Execution Accuracy": 0.61
24
+ },
25
+ "spider_syn": {
26
+ "Exact Match": 0.408,
27
+ "Execution Accuracy": 0.467
28
+ },
29
+ "spider_realistic": {
30
+ "Exact Match": 0.457,
31
+ "Execution Accuracy": 0.502
32
+ },
33
+ "spider_dk": {
34
+ "Exact Match": 0.406,
35
+ "Execution Accuracy": 0.452
36
+ }
37
+ },
38
+ "artifacts": {
39
+ "spider_data": {
40
+ "run_name": "baselines__train_rkl__ckpt872__test",
41
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
42
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
43
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
44
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
45
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
46
+ "ckpt_step": "872"
47
+ },
48
+ "spider_syn": {
49
+ "run_name": "baselines__train_rkl__ckpt872__test",
50
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
51
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
52
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
53
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
54
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
55
+ "ckpt_step": "872"
56
+ },
57
+ "spider_realistic": {
58
+ "run_name": "baselines__train_rkl__ckpt872__test",
59
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
60
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
61
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
62
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
63
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
64
+ "ckpt_step": "872"
65
+ },
66
+ "spider_dk": {
67
+ "run_name": "baselines__train_rkl__ckpt872__test",
68
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
69
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
70
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
71
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
72
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
73
+ "ckpt_step": "872"
74
+ }
75
+ }
76
+ }
77
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/seed50/eval_results.json ADDED
@@ -0,0 +1,514 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "seed": 50,
3
+ "generated_at": "2026-07-20T12:37:18.612447+00:00",
4
+ "selection": {
5
+ "primary_metric": "Execution Accuracy",
6
+ "tie_breaker": "Exact Match"
7
+ },
8
+ "results": [
9
+ {
10
+ "run_name": "baselines__train_amid__ckpt872__test",
11
+ "benchmark": "spider_data",
12
+ "seed": 50,
13
+ "scores": {
14
+ "Exact Match": 0.51,
15
+ "Execution Accuracy": 0.572
16
+ },
17
+ "artifacts": {
18
+ "run_name": "baselines__train_amid__ckpt872__test",
19
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
20
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
21
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
22
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
23
+ "train_script": "scripts/llama/baselines/train_amid.sh",
24
+ "ckpt_step": "872"
25
+ }
26
+ },
27
+ {
28
+ "run_name": "baselines__train_amid__ckpt872__test",
29
+ "benchmark": "spider_dk",
30
+ "seed": 50,
31
+ "scores": {
32
+ "Exact Match": 0.391,
33
+ "Execution Accuracy": 0.471
34
+ },
35
+ "artifacts": {
36
+ "run_name": "baselines__train_amid__ckpt872__test",
37
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
38
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
39
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
40
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
41
+ "train_script": "scripts/llama/baselines/train_amid.sh",
42
+ "ckpt_step": "872"
43
+ }
44
+ },
45
+ {
46
+ "run_name": "baselines__train_amid__ckpt872__test",
47
+ "benchmark": "spider_realistic",
48
+ "seed": 50,
49
+ "scores": {
50
+ "Exact Match": 0.415,
51
+ "Execution Accuracy": 0.47
52
+ },
53
+ "artifacts": {
54
+ "run_name": "baselines__train_amid__ckpt872__test",
55
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
56
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
57
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
58
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
59
+ "train_script": "scripts/llama/baselines/train_amid.sh",
60
+ "ckpt_step": "872"
61
+ }
62
+ },
63
+ {
64
+ "run_name": "baselines__train_amid__ckpt872__test",
65
+ "benchmark": "spider_syn",
66
+ "seed": 50,
67
+ "scores": {
68
+ "Exact Match": 0.354,
69
+ "Execution Accuracy": 0.422
70
+ },
71
+ "artifacts": {
72
+ "run_name": "baselines__train_amid__ckpt872__test",
73
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_amid__ckpt872__test.pred.sql",
74
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_amid__ckpt872__test.gold.sql",
75
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log",
76
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_amid_spider_amid_adaptive-amid_ab_pr_0.5_0.5_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
77
+ "train_script": "scripts/llama/baselines/train_amid.sh",
78
+ "ckpt_step": "872"
79
+ }
80
+ },
81
+ {
82
+ "run_name": "baselines__train_csd__ckpt654__test",
83
+ "benchmark": "spider_data",
84
+ "seed": 50,
85
+ "scores": {
86
+ "Exact Match": 0.538,
87
+ "Execution Accuracy": 0.599
88
+ },
89
+ "artifacts": {
90
+ "run_name": "baselines__train_csd__ckpt654__test",
91
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
92
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
93
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
94
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
95
+ "train_script": "scripts/llama/baselines/train_csd.sh",
96
+ "ckpt_step": "654"
97
+ }
98
+ },
99
+ {
100
+ "run_name": "baselines__train_csd__ckpt654__test",
101
+ "benchmark": "spider_dk",
102
+ "seed": 50,
103
+ "scores": {
104
+ "Exact Match": 0.419,
105
+ "Execution Accuracy": 0.469
106
+ },
107
+ "artifacts": {
108
+ "run_name": "baselines__train_csd__ckpt654__test",
109
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
110
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
111
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
112
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
113
+ "train_script": "scripts/llama/baselines/train_csd.sh",
114
+ "ckpt_step": "654"
115
+ }
116
+ },
117
+ {
118
+ "run_name": "baselines__train_csd__ckpt654__test",
119
+ "benchmark": "spider_realistic",
120
+ "seed": 50,
121
+ "scores": {
122
+ "Exact Match": 0.445,
123
+ "Execution Accuracy": 0.486
124
+ },
125
+ "artifacts": {
126
+ "run_name": "baselines__train_csd__ckpt654__test",
127
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
128
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
129
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
130
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
131
+ "train_script": "scripts/llama/baselines/train_csd.sh",
132
+ "ckpt_step": "654"
133
+ }
134
+ },
135
+ {
136
+ "run_name": "baselines__train_csd__ckpt654__test",
137
+ "benchmark": "spider_syn",
138
+ "seed": 50,
139
+ "scores": {
140
+ "Exact Match": 0.414,
141
+ "Execution Accuracy": 0.471
142
+ },
143
+ "artifacts": {
144
+ "run_name": "baselines__train_csd__ckpt654__test",
145
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_csd__ckpt654__test.pred.sql",
146
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_csd__ckpt654__test.gold.sql",
147
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log",
148
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_csd_spider_csd_csd_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
149
+ "train_script": "scripts/llama/baselines/train_csd.sh",
150
+ "ckpt_step": "654"
151
+ }
152
+ },
153
+ {
154
+ "run_name": "baselines__train_distillm__ckpt654__test",
155
+ "benchmark": "spider_data",
156
+ "seed": 50,
157
+ "scores": {
158
+ "Exact Match": 0.504,
159
+ "Execution Accuracy": 0.58
160
+ },
161
+ "artifacts": {
162
+ "run_name": "baselines__train_distillm__ckpt654__test",
163
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
164
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
165
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
166
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
167
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
168
+ "ckpt_step": "654"
169
+ }
170
+ },
171
+ {
172
+ "run_name": "baselines__train_distillm__ckpt654__test",
173
+ "benchmark": "spider_dk",
174
+ "seed": 50,
175
+ "scores": {
176
+ "Exact Match": 0.398,
177
+ "Execution Accuracy": 0.45
178
+ },
179
+ "artifacts": {
180
+ "run_name": "baselines__train_distillm__ckpt654__test",
181
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
182
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
183
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
184
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
185
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
186
+ "ckpt_step": "654"
187
+ }
188
+ },
189
+ {
190
+ "run_name": "baselines__train_distillm__ckpt654__test",
191
+ "benchmark": "spider_realistic",
192
+ "seed": 50,
193
+ "scores": {
194
+ "Exact Match": 0.423,
195
+ "Execution Accuracy": 0.474
196
+ },
197
+ "artifacts": {
198
+ "run_name": "baselines__train_distillm__ckpt654__test",
199
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
200
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
201
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
202
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
203
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
204
+ "ckpt_step": "654"
205
+ }
206
+ },
207
+ {
208
+ "run_name": "baselines__train_distillm__ckpt654__test",
209
+ "benchmark": "spider_syn",
210
+ "seed": 50,
211
+ "scores": {
212
+ "Exact Match": 0.368,
213
+ "Execution Accuracy": 0.432
214
+ },
215
+ "artifacts": {
216
+ "run_name": "baselines__train_distillm__ckpt654__test",
217
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_distillm__ckpt654__test.pred.sql",
218
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_distillm__ckpt654__test.gold.sql",
219
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log",
220
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_distillm_spider_distillm_adaptive-srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/654",
221
+ "train_script": "scripts/llama/baselines/train_distillm.sh",
222
+ "ckpt_step": "654"
223
+ }
224
+ },
225
+ {
226
+ "run_name": "baselines__train_fkl__ckpt872__test",
227
+ "benchmark": "spider_data",
228
+ "seed": 50,
229
+ "scores": {
230
+ "Exact Match": 0.511,
231
+ "Execution Accuracy": 0.57
232
+ },
233
+ "artifacts": {
234
+ "run_name": "baselines__train_fkl__ckpt872__test",
235
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
236
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
237
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
238
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
239
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
240
+ "ckpt_step": "872"
241
+ }
242
+ },
243
+ {
244
+ "run_name": "baselines__train_fkl__ckpt872__test",
245
+ "benchmark": "spider_dk",
246
+ "seed": 50,
247
+ "scores": {
248
+ "Exact Match": 0.381,
249
+ "Execution Accuracy": 0.445
250
+ },
251
+ "artifacts": {
252
+ "run_name": "baselines__train_fkl__ckpt872__test",
253
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
254
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
255
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
256
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
257
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
258
+ "ckpt_step": "872"
259
+ }
260
+ },
261
+ {
262
+ "run_name": "baselines__train_fkl__ckpt872__test",
263
+ "benchmark": "spider_realistic",
264
+ "seed": 50,
265
+ "scores": {
266
+ "Exact Match": 0.415,
267
+ "Execution Accuracy": 0.465
268
+ },
269
+ "artifacts": {
270
+ "run_name": "baselines__train_fkl__ckpt872__test",
271
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
272
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
273
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
274
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
275
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
276
+ "ckpt_step": "872"
277
+ }
278
+ },
279
+ {
280
+ "run_name": "baselines__train_fkl__ckpt872__test",
281
+ "benchmark": "spider_syn",
282
+ "seed": 50,
283
+ "scores": {
284
+ "Exact Match": 0.363,
285
+ "Execution Accuracy": 0.435
286
+ },
287
+ "artifacts": {
288
+ "run_name": "baselines__train_fkl__ckpt872__test",
289
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_fkl__ckpt872__test.pred.sql",
290
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_fkl__ckpt872__test.gold.sql",
291
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log",
292
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_fkl_spider_fkl_fkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
293
+ "train_script": "scripts/llama/baselines/train_fkl.sh",
294
+ "ckpt_step": "872"
295
+ }
296
+ },
297
+ {
298
+ "run_name": "baselines__train_rkl__ckpt872__test",
299
+ "benchmark": "spider_data",
300
+ "seed": 50,
301
+ "scores": {
302
+ "Exact Match": 0.547,
303
+ "Execution Accuracy": 0.61
304
+ },
305
+ "artifacts": {
306
+ "run_name": "baselines__train_rkl__ckpt872__test",
307
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
308
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
309
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
310
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
311
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
312
+ "ckpt_step": "872"
313
+ }
314
+ },
315
+ {
316
+ "run_name": "baselines__train_rkl__ckpt872__test",
317
+ "benchmark": "spider_dk",
318
+ "seed": 50,
319
+ "scores": {
320
+ "Exact Match": 0.406,
321
+ "Execution Accuracy": 0.452
322
+ },
323
+ "artifacts": {
324
+ "run_name": "baselines__train_rkl__ckpt872__test",
325
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
326
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
327
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
328
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
329
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
330
+ "ckpt_step": "872"
331
+ }
332
+ },
333
+ {
334
+ "run_name": "baselines__train_rkl__ckpt872__test",
335
+ "benchmark": "spider_realistic",
336
+ "seed": 50,
337
+ "scores": {
338
+ "Exact Match": 0.457,
339
+ "Execution Accuracy": 0.502
340
+ },
341
+ "artifacts": {
342
+ "run_name": "baselines__train_rkl__ckpt872__test",
343
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
344
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
345
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
346
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
347
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
348
+ "ckpt_step": "872"
349
+ }
350
+ },
351
+ {
352
+ "run_name": "baselines__train_rkl__ckpt872__test",
353
+ "benchmark": "spider_syn",
354
+ "seed": 50,
355
+ "scores": {
356
+ "Exact Match": 0.408,
357
+ "Execution Accuracy": 0.467
358
+ },
359
+ "artifacts": {
360
+ "run_name": "baselines__train_rkl__ckpt872__test",
361
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_rkl__ckpt872__test.pred.sql",
362
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_rkl__ckpt872__test.gold.sql",
363
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log",
364
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_rkl_spider_rkl_rkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/872",
365
+ "train_script": "scripts/llama/baselines/train_rkl.sh",
366
+ "ckpt_step": "872"
367
+ }
368
+ },
369
+ {
370
+ "run_name": "baselines__train_sfkl__ckpt654__test",
371
+ "benchmark": "spider_data",
372
+ "seed": 50,
373
+ "scores": {
374
+ "Exact Match": 0.5,
375
+ "Execution Accuracy": 0.571
376
+ },
377
+ "artifacts": {
378
+ "run_name": "baselines__train_sfkl__ckpt654__test",
379
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
380
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
381
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
382
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
383
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
384
+ "ckpt_step": "654"
385
+ }
386
+ },
387
+ {
388
+ "run_name": "baselines__train_sfkl__ckpt654__test",
389
+ "benchmark": "spider_dk",
390
+ "seed": 50,
391
+ "scores": {
392
+ "Exact Match": 0.396,
393
+ "Execution Accuracy": 0.46
394
+ },
395
+ "artifacts": {
396
+ "run_name": "baselines__train_sfkl__ckpt654__test",
397
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
398
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
399
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
400
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
401
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
402
+ "ckpt_step": "654"
403
+ }
404
+ },
405
+ {
406
+ "run_name": "baselines__train_sfkl__ckpt654__test",
407
+ "benchmark": "spider_realistic",
408
+ "seed": 50,
409
+ "scores": {
410
+ "Exact Match": 0.439,
411
+ "Execution Accuracy": 0.48
412
+ },
413
+ "artifacts": {
414
+ "run_name": "baselines__train_sfkl__ckpt654__test",
415
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
416
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
417
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
418
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
419
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
420
+ "ckpt_step": "654"
421
+ }
422
+ },
423
+ {
424
+ "run_name": "baselines__train_sfkl__ckpt654__test",
425
+ "benchmark": "spider_syn",
426
+ "seed": 50,
427
+ "scores": {
428
+ "Exact Match": 0.384,
429
+ "Execution Accuracy": 0.432
430
+ },
431
+ "artifacts": {
432
+ "run_name": "baselines__train_sfkl__ckpt654__test",
433
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.pred.sql",
434
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_sfkl__ckpt654__test.gold.sql",
435
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log",
436
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_sfkl_spider_sfkl_sfkl_e5-bs2-lr0.0001-G8-N2-NN1-kd0.7-lora-16-64-0.1/654",
437
+ "train_script": "scripts/llama/baselines/train_sfkl.sh",
438
+ "ckpt_step": "654"
439
+ }
440
+ },
441
+ {
442
+ "run_name": "baselines__train_srkl__ckpt436__test",
443
+ "benchmark": "spider_data",
444
+ "seed": 50,
445
+ "scores": {
446
+ "Exact Match": 0.482,
447
+ "Execution Accuracy": 0.555
448
+ },
449
+ "artifacts": {
450
+ "run_name": "baselines__train_srkl__ckpt436__test",
451
+ "pred_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
452
+ "gold_path": "results/infer/llama_baselines/llama/spider_data/seed50/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
453
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
454
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
455
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
456
+ "ckpt_step": "436"
457
+ }
458
+ },
459
+ {
460
+ "run_name": "baselines__train_srkl__ckpt436__test",
461
+ "benchmark": "spider_dk",
462
+ "seed": 50,
463
+ "scores": {
464
+ "Exact Match": 0.342,
465
+ "Execution Accuracy": 0.415
466
+ },
467
+ "artifacts": {
468
+ "run_name": "baselines__train_srkl__ckpt436__test",
469
+ "pred_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
470
+ "gold_path": "results/infer/llama_baselines/llama/spider_dk/seed50/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
471
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_dk/seed50/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
472
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
473
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
474
+ "ckpt_step": "436"
475
+ }
476
+ },
477
+ {
478
+ "run_name": "baselines__train_srkl__ckpt436__test",
479
+ "benchmark": "spider_realistic",
480
+ "seed": 50,
481
+ "scores": {
482
+ "Exact Match": 0.39,
483
+ "Execution Accuracy": 0.451
484
+ },
485
+ "artifacts": {
486
+ "run_name": "baselines__train_srkl__ckpt436__test",
487
+ "pred_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
488
+ "gold_path": "results/infer/llama_baselines/llama/spider_realistic/seed50/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
489
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_realistic/seed50/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
490
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
491
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
492
+ "ckpt_step": "436"
493
+ }
494
+ },
495
+ {
496
+ "run_name": "baselines__train_srkl__ckpt436__test",
497
+ "benchmark": "spider_syn",
498
+ "seed": 50,
499
+ "scores": {
500
+ "Exact Match": 0.355,
501
+ "Execution Accuracy": 0.411
502
+ },
503
+ "artifacts": {
504
+ "run_name": "baselines__train_srkl__ckpt436__test",
505
+ "pred_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_srkl__ckpt436__test.pred.sql",
506
+ "gold_path": "results/infer/llama_baselines/llama/spider_syn/seed50/formatted_data/baselines__train_srkl__ckpt436__test.gold.sql",
507
+ "eval_log_path": "results/eval/llama_baselines/llama/spider_syn/seed50/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log",
508
+ "ckpt_path": "/mnt/hungpv/projects/text2sql_distillation_draft/results/llama_baselines/baselines_train_srkl_spider_srkl_srkl_e5-bs1-lr0.0001-G16-N2-NN1-kd0.7-lora-16-64-0.1/436",
509
+ "train_script": "scripts/llama/baselines/train_srkl.sh",
510
+ "ckpt_step": "436"
511
+ }
512
+ }
513
+ ]
514
+ }
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed10/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed100/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed1234/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed42/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_amid__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_csd__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_distillm__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_fkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_rkl__ckpt872__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_sfkl__ckpt654__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff
 
text2sql_distillation_draft/results/eval/llama_baselines/llama/spider_data/seed50/baselines__train_srkl__ckpt436__test.etype-all.timeout-60.log ADDED
The diff for this file is too large to render. See raw diff