iamPi/sabaton-914184bb

sabaton-914184bb

Cleaned SFT corpus for Bittensor SN120 (Affine), Reason v4 (weight_version_key=7). 17,655 rows, TRL prompt-completion format.

Format is exactly `</think> THOUGHT: {z}

{y}` on every row.

How it was cleaned

Derived from `iamPi/tessera-77d11909`. Prompts and actions are byte-identical to the source; only the thought channel changed, and 483 rows (2.66%) were removed.

Repaired in place (40.08% of rows had at least one defect)

The retag matters: the validator's BASH_RE matches only `bash , and split_rollout takes the last match as the action. A thought that quotes a bash fence could therefore be mistaken for the command whenever a model's real fence came out malformed.

Removed (483 rows)

Those last five were collapsed multi-turn transcripts: the "thought" contained the result of a command that had not been run yet, which would train the model to hallucinate observations.

Two-channel thoughts — latent analysis followed by a visible summary — were kept. That is the contract's own shape: split_rollout joins the latent and visible halves on purpose, and 57 of the 61 such rows carry genuinely distinct text in each half.

Guarantees

Every row was checked after rewriting: it round-trips through the validator's own split_rollout, is byte-identical to the force_text() rendering, has no think tag / THOUGHT label / bash fence / harness tag in its thought, clears the 80-char floor, and fits the rollout budget.

Cue

A fixed cue ends every thought:

</think>
THOUGHT: {z}

Next bash command to run now:

{y}

The `

before the action is kept because inject_prompt() hardcodes that separator, so the trained string is byte-identical to what the teacher scores. 14 rows trip the fuzzy leakage heuristic because the cue contributes the tokens bash / command / run` to short commands that literally invoke bash.

The row set is identical to the no-cue build, so the two are a clean A/B.