iamPi/halyard-9b1a09e7

halyard-9b1a09e7

Format-repaired SFT corpus for Bittensor SN120 (Affine), Reason v4 (weight_version_key=7). TRL prompt-completion format.

What was repaired

Every row's thought channel was normalised to exactly `</think> THOUGHT: {{z}}

{{y}}`. 40.08% of rows (7,270 / 18,138) carried at least one defect:

Nothing was dropped. Tags and labels are markers, so removing them leaves the reasoning intact. 61 rows carry up to 8,077 characters of real analysis before their second label, so a "keep only what follows the last THOUGHT:" rule would have deleted it.

Row set, prompts and actions are byte-identical to `iamPi/tessera-77d11909`; only the thought text changed. Every row was re-parsed with the validator's own split_rollout() and checked byte-for-byte against the force_text() rendering before writing.

Cue

A fixed cue is appended to the end of every thought:

</think>
THOUGHT: {z}

Next bash command to run now:

{y}

The `

before the action is kept because the validator's inject_prompt()` hardcodes that separator, so this is the only construction where the trained string is byte-identical to the string the teacher scores.

Known caveats

• —1 row(s) fall below the 80-character crown floor. Their length was previously made up partly by the junk that has now been removed.
• —14 row(s) trip the validator's fuzzy leakage heuristic (the cue contributes the tokens bash/command/run, which tips short commands that literally invoke bash).
• —35 row(s) exceed the 1,792-token rollout budget — pre-existing in tessera, not introduced here.

Built with scripts/normalize_sft_format.py.