AlexWortega/flint-opus-curriculum-20261026

Flint Opus-judged curriculum RL, 2026-10-26

Final archive. Latest observed cumulative optimizer step: 200/200. Root adapter checkpoint: 200.

Fresh attention-only LoRA r16/alpha32 on the base model. Curriculum: steps 1-100 antislop only; steps 101-200 character-disjoint roleplay and real Pi/Harbor episodes. Text rewards use the astra4 calibrated-v1 unslop formula graded by the Opus 5.5 judge it was calibrated on (no cross-judge mapping); roleplay mu/sigma are fit on base-model outputs at step 0. Reasoning slop is scored by Opus and subtracted per episode. LR 1e-5, KL beta .04. Trainer/vLLM kernel differences use TRL-style token-level importance correction capped at 2; this is not exact numeric parity.

Base model is pinned at 2a247be0c0aec895ccb0ac7ca92229ec6df37288. The root adapter is the latest archived checkpoint, not a claim of the best checkpoint. The base model weights remain available from the pinned upstream repository.

code/ preserves the updated source, configurations, data splits and provenance. Final archives include initial/continuation checkpoints, optimizer states, native token/logprob receipts, judge outputs and costs, terminal trials and logs, and evaluation artifacts. All checkpoints are also under checkpoints/. Credentials are excluded/redacted. Full original raw exports remain on the CPU host.

This experiment explicitly trains on three Terminal-Bench tasks; training success is not a benchmark score. Roleplay grading abstentions are excluded and counted; truncated text outputs receive zero reward. Eval denominators and training/heldout splits are saved. No convergence or quality improvement is established by training completion. Source datasets and model retain their respective upstream terms.

Artifact file sizes and hashes are recorded in artifact_manifest.json and checked against the immutable Hugging Face commit before infrastructure deletion.