chimera evolve

Opt-in model evolution (curate trajectories -> LoRA/DPO recipe).

Subcommands

chimera evolve export

Export a curated SFT or DPO dataset from trajectories.

Options

--outstrrequired
Output JSONL path.
--formatstrdefault: 'sft'
sft | dpo
--trajstr
Trajectory JSONL (default: <home>/trajectories.jsonl).
--min-rewardfloatdefault: 0.0
Drop examples below this reward.
--no-dedupboolean
Keep duplicate examples.
--min-marginfloatdefault: 0.0
DPO: min reward margin chosen − rejected.
--min-stepsintdefault: 0
Recipe: keep only traces with >= N steps.
--diverseboolean
Recipe: at most one SFT example per task.
--min-processfloat
Keep only traces whose step-following score >= this (SkillCoach).

chimera evolve guard

Watch evolution health; retract the most recent skill on a SIGNIFICANT regression (M19-A6).

Options

--limitintdefault: 0
Limit demo tasks (0 = all).
--model, -mstr
Override the model slug.
--cost-drift-tolfloat
Also roll back if second-half mean cost exceeds first by more than this.
--applyboolean
Retire the most recent skill IF a SIGNIFICANT regression is measured.

chimera evolve recipe

Emit a runnable LoRA training recipe (train.py + README + requirements).

Options

--outstrrequired
Directory for the training recipe.
--formatstrdefault: 'sft'
sft | dpo
--base-modelstrdefault: 'meta-llama/Llama-3.1-8B-Instruct'
--datasetstrdefault: 'dataset.jsonl'
Dataset filename the script reads.

chimera evolve refine

GEPA-refine a skill from verified trajectories, gated on non-regressing transfer (M19-A5).

Arguments

SKILL_NAMEstrrequired
Name of the learned skill to refine.

Options

--trajstr
Trajectory JSONL (default: <home>/trajectories.jsonl).
--model, -mstr
Override the model.
--budgetintdefault: 20
GEPA rollout budget.
--min-rewardfloatdefault: 1.0
Only mine trajectories at/above this reward (1.0 = verified).
--applyboolean
Persist the refined skill IF it passes the transfer gate.

chimera evolve rft

One rejection-sampling fine-tuning round, gated by an honest A/B on two bench result files. Rejection-samples the collected trajectories (successes at/above the reward bar), then promotes the round ONLY if the candidate beats the baseline with a confidence interval that excludes zero — no lift, no promotion, no training on noise. Artifacts are withheld for an unpromoted round unless ``--force``. Feed ``--baseline``/``--candidate`` the pass/fail lists two bench runs produce.

Options

--baselinestrrequired
JSON list of baseline bench pass/fail.
--candidatestrrequired
JSON list of candidate bench pass/fail.
--trajstr
Trajectory JSONL (default: <home>/trajectories.jsonl).
--min-rewardfloatdefault: 0.5
Rejection-sampling reward bar.
--min-examplesintdefault: 30
Accepted examples needed to gate.
--top-kintdefault: 0
Keep at most this many accepted per prompt (0 = all).
--outstr
If promoted, write dataset + recipe here.
--forceboolean
Export even if the round is not promoted.

chimera evolve status

Show how much training signal the collected trajectories hold.

Options

--trajstr
Trajectory JSONL (default: <home>/trajectories.jsonl).
--min-rewardfloatdefault: 0.0
Drop examples below this reward.
--min-examplesintdefault: 30
Examples needed before training is worth it.

chimera evolve tune

Self-optimize the agent spec (OpenJarvis meta-search) against the daily scenarios. Each round a model proposes a coordinated edit to the spec; the candidate is scored on the daily scenarios and kept only on non-regression. Uses real model calls.

Options

--roundsintdefault: 2
Meta-search rounds.
--modelstr
Base model for the spec.
--max-stepsintdefault: 8
Initial runtime step budget.