jkminder/model-raising-pbsft-eval

Tasks

• text-generation

Citation text from the card

The frozen Qwen generator makes citation errors — over-citation (decorative `[X.Y]`),
wrong-section ids, and missed citations. To provide a **corrected citation reference**, every
row was independently audited by a Claude agent against the charter and the v11 citation rubric
(subtractive test for load-bearingness, the v11 cheatsheet for the correct section id, default
1 / max 2 cites). The audit fixes the `[X.Y]` brackets (citations-first, with light latitude to
repair a wrong response); the result is `claude_cleaned`.

| Column | Type | Description |
|---|---|---|
| `claude_cleaned` | string | The audited/corrected `cited` response, with corrected `[X.Y]` markers. |
| `claude_final_citations` | list[string] | The `[X.Y]` ids present in `claude_cleaned`. |
| `claude_changed` | bool | Whether `claude_cleaned` differs from the original `cited`. |
| `claude_action` | string | `kept` / `removed_decorative` / `retargeted` / `added` / `mixed` / `rewrote` / `blocked`. |
| `claude_citations_correct_before` | bool | Whether the original citations were all correct **and** sufficient. |
| `claude_response_quality_ok` | bool | Whether the underlying response was sound (audit flag; prose not rewritten unless `action=rewrote`). |
| `claude_reason` | string | Short rationale for the edit/keep. |
| `claude_confidence` | string | `high` / `med` / `low` (`n/a` for blocked rows). |
| `claude_model` | string | Which model produced the result: `opus` / `sonnet` / `deterministic` / `blocked`. |
| 

All datasets marked Permissive or share-alike | JSON