File size: 1,343 Bytes
ad1725b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
{
  "type": "code2lora_gru_commit_smartcap",
  "in_dir": "/scratch/lhotsko/REPO_DATASET/commit_parquet_hf",
  "max_per_file": 4,
  "max_per_commit": 8,
  "min_target_chars": 4,
  "seed": 3407,
  "splits_capped": [
    "train"
  ],
  "splits_passthrough": [
    "cr_val",
    "cr_test"
  ],
  "per_split": {
    "train": {
      "qna_path_in": "/scratch/lhotsko/REPO_DATASET/commit_parquet_hf/qna/train.parquet",
      "qna_path_out": "/scratch/lhotsko/REPO_DATASET/commit_parquet_hf_smartcap/qna/train.parquet",
      "rows_in_total": 3639660,
      "rows_in_train": 2667430,
      "rows_in_eval": 972230,
      "rows_out_train_capped": 215129,
      "rows_out_eval_passthrough": 972230,
      "rows_dropped_trivial_target": 550621,
      "rows_dropped_smart_cap": 1901680,
      "n_repos_seen": 400,
      "n_repos_with_train_kept": 400,
      "n_commits_with_train_kept": 45516,
      "max_per_file": 4,
      "max_per_commit": 8,
      "min_target_chars": 4,
      "seed": 3407,
      "elapsed_sec": 212.08423447608948
    }
  },
  "notes": "Only in_repo_split=='train' rows are capped. val / test rows are passed through verbatim. Quality filter drops trivial targets (length < min_target_chars or bare ')' / leading-comma tokens). Smart cap interleaves rows across (test_file, test_function) groups before truncating to max_per_commit."
}