iamPi/tessera-0f0e8791

tessera (cleaned)

Format-normalized derivative of `iamPi/tessera-77d11909`. Same rows, same prompts; the completions have been repaired to the Affine SWE turn contract and given a fixed action cue.

Completion format

Every completion is exactly:

</think>
THOUGHT: {thought}

The analysis is complete. Next command:

{command}

The prompt ends inside an open <think> block, so the completion opens with </think>. thought is plain prose — no tags, no second label, no fences of its own — and the action is a single closed ``bash block at the very end, with nothing after it. This round-trips exactly through the validator's own split_rollout`.

What changed vs. the source

18,138 rows in, 18,134 out (4 dropped: 3 empty thoughts, 1 with a hallucinated environment turn). 7,396 rows had at least one format repair:

All 18,134 completions then received the \n\nThe analysis is complete. Next command: cue between the thought and the action. This is the exact suffix the reigning king (vera6/affine-5g4yy75zuz-t6, reign 36) emits on 99.4% of its own duel thoughts.

A sibling build carrying the cue \n\nNext bash command to run now: instead lives at `iamPi/tessera-67843aff`. The two datasets are byte-identical apart from that one string.

Extra columns

• —thought — the bare reasoning, without the cue (use this for length filtering and dedup; the cue is present only in completion).
• —action — the single ```bash block.
• —clean_flags — comma-separated repairs applied to that row, including thought_below_gate for the 444 rows whose bare thought is under the 80-char min_thought_chars gate. The gate is measured before the cue is appended, so the boilerplate does not mask a thin thought.

Known characteristics (not repaired)

• —409 rows have negative teacher_lift; 3,079 have n_pool_refs == 1, where modal_share == 1.0 is vacuous.
• —401 duplicate completions, 367 of which are one turn-0 opener paired with ls -la.
• —phase is empty on 1,069 source == 'swe' rows.
• —Estimated total sequence length: median ~5.7k tokens, 13% over 16k, max 31k.