---
language:
- en
license: other
---

# Missing-answer teacher reruns

Source: [JonesLin/multi-model-cot-missing-answers](https://huggingface.co/datasets/JonesLin/multi-model-cot-missing-answers), pinned at commit
`f32e2f1fc0bcdfd7e0b61929cc494126526e09c9`. Open-ended rows were excluded, then rows were deduplicated by
`stage1_id` within the original train and validation splits. The source has 10,945
eligible questions (10,715 train; 230 validation).

This dataset contains one Luna response per completed question. Requests use
`rerun-yibingwei/gpt-6-luna` and `rerun-jiang/gpt-6-luna` alternately; both
resolve to `gpt-6-luna` on CLI Proxy 7.3.17. The routed model ID used for each
request is recorded in `generator`.

When a Luna reply still fails the format checks after three replies, the same
question is retried with `gpt-6-sol` through the two prefixed Codex accounts.
Those records appear in `data/sol_fallback.jsonl`, with the original Luna
failure preserved in `data/luna.jsonl` and its reason copied into
`luna_failure_reason` on the fallback row. The fallback model ID is recorded
per row in `generator`.

System message, verbatim:

```text
You are a careful expert problem solver.
```

User message template, verbatim (`{prompt}` is replaced with the original prompt):

```text
{prompt}

Solve the problem above. Write your reasoning step by step in your reply; the written reasoning is what we keep, so do not rely on hidden reasoning. Keep it concise but complete.
End your reply with exactly one final line in this form:
Final answer: <answer>
Rules for <answer>:
- If the question lists labelled choices, give only the label of one listed choice, for example "Final answer: C".
- For math or numeric questions, give only the final value or expression.
- Otherwise, give the shortest complete answer.
Always commit to one answer, even if unsure. Do not ask for more information.
```

Generation settings: model defaults for sampling (no temperature, top-p, or
similar parameters); `max_tokens=16384`; 600-second request timeout. Full-run
concurrency: Luna 32 total across the two Codex accounts. Failed format checks
are retried for up to three model replies. HTTP 429 and 5xx errors use
exponential backoff. `cot` stores only the
assistant reply body; provider `reasoning_content` is a separate nullable field.
The scheduler keeps 32 CLI Proxy work batches active, each with up to eight
questions processed as independent model requests, and replaces a batch when
it finishes.

Run timestamps (UTC): 2026-09-27T01:07:05.345749+00:00 to 2026-09-27T02:18:43.651536+00:00. Snapshot status: final.

| File | Rows | Success | Failed |
| --- | ---: | ---: | ---: |
| `data/luna.jsonl` | 10945 | 10932 | 13 |
| `data/sol_fallback.jsonl` | 13 | 13 | 0 |

Each row includes the source question metadata, source response IDs and failed
teachers, the generated response and final answer, usage, validation outcome,
and creation time. `failed` rows preserve the final raw reply and failure reason.