---
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl
      - split: test
        path: data/test.jsonl
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
size_categories:
  - 1K<n<10K
---

# jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50-primecompat

Prime `prime-rl` supervised fine-tuning dataset for Humanize-RL.

This is the `env0315_clean50` S2 repair-data candidate. It starts from the
env0314 Prime SFT corpus and adds cleaned env0315 repair references generated
from saved Prime rollout-audit failures.

## Splits

| split | rows |
|---|---:|
| train | 4358 |
| validation | 242 |
| test | 243 |
| total | 4843 |

## Sources

| source | rows |
|---|---:|
| safe_expand_3000_raw | 1801 |
| stream_b | 1556 |
| safe_expand_raw | 667 |
| chat_expanded | 423 |
| unknown | 326 |
| prime_failure_reference_generation | 70 |

Accepted repair-reference rows: `70`.

## Modes

| mode | rows |
|---|---:|
| rewrite_humanize | 2999 |
| direct_generation | 1803 |
| rewrite | 20 |
| tone_shift | 7 |
| long_form_generate | 7 |
| compression | 4 |
| multi_constraint_compose | 2 |
| expansion | 1 |

## Schema

Training rows are restricted to Prime-compatible feature types. Each row includes
a `messages` column with exactly two turns:

```json
[
  {"role": "user", "content": "..."},
  {"role": "assistant", "content": "..."}
]
```

Nested audit fields such as `quality` and `metadata` are kept in the local
builder reports, not in the Hub training rows, because Prime's pinned
`datasets` stack does not accept Hub features exported as `_type: Json`.

## Intended Use

Use this dataset for the S2 Qwen 2B Prime dataset-SFT ablation. Do not treat it
as a promoted model result by itself. SFT output still has to beat base on the
frozen eval prompts, pass the detector-mimic gate, pass human read, and pass the
SFT promotion gate before any RL-after-SFT launch.