---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl.gz
  - split: dev
    path: data/dev.jsonl.gz
  - split: test
    path: data/test.jsonl.gz
---

# SWE-Fixer Diff Train Full, Repo-Disjoint Qwen3 Splits

This dataset is derived from `.cache/glotcond/materialized/swe_fixer_diff_train_full.jsonl`.
It adds `prompt_input_token_count`, computed with `Qwen/Qwen3-4B-Base` using
`add_special_tokens=False`.

The split is repo-disjoint by `metadata.repo` and targets train/dev/test ratios
of 0.7/0.1/0.2 while balancing the prompt token-count distribution.

Column order: `id`, `prompt`, `answer`, `prompt_input_token_count`, `metadata`.

| split | rows | row ratio | repos | token mean | median | p90 | p95 | max |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| train | 33387 | 0.7000 | 240 | 31968.8 | 15708.0 | 76027.0 | 113771.6 | 1135223 |
| dev | 4778 | 0.1002 | 33 | 31899.0 | 15805.5 | 79846.3 | 129640.6 | 450936 |
| test | 9534 | 0.1999 | 70 | 32040.3 | 15861.0 | 76438.9 | 112687.2 | 483488 |

`metadata/split_summary.json` contains the full split diagnostics.