# LoRA Training History

Saved adapter-checkpoint lineage, training configuration, dataset growth, stability diagnostics, and reward-consumption history for the current fine-tuned model runs.

- Saved checkpoints summarized: 7
- Checkpoint span: March 3, 2026 to March 19, 2026
- Base model: unsloth/gpt-oss-20b-unsloth-bnb-4bit
- Training stack: UnSloth + PyTorch 2.9.1+cu128 + Transformers 4.57.1 + PEFT 0.18.0
- LoRA rank / alpha: 8 / 16
- Latest run training set: 168 examples (1824 tracked definitions)
- Archive state at latest checkpoint: 32 successful non-replace observations overall
- Current staged corpus: 951 SFT pairs
- Platform: ml.g6e.xlarge in us-east-2
- Sequence limits: 4600 training tokens, 600 max completion tokens

Training and serving setup

- Fine-tuning method: Reward-weighted SFT on credited llm_output artifacts from successful observations
- Serving model: Base model plus the latest LoRA adapter checkpoint
- Training stabilizer: clip_higher (clip_higher ratio clipping against cached completion log-probabilities to limit destabilizing policy jumps)
- Training precision / optimizer: bf16 with adamw_8bit
- Recipe tuning across runs: We tuned some training-job parameters across runs while keeping the core setup fixed. Learning rate was 5e-05 through v6 and changed to 4e-05 at v7. A 168-example training cap was introduced at v4. Core model, LoRA shape, optimizer, precision, and hardware remained the same.
- Checkpoint progression: Each saved adapter resumes from the previous saved adapter when available
- Target modules adapted: 60
- Included raw records are reader-facing summaries of the latest checkpoint provenance, latest run metrics, and latest reward-consumption receipt, with host-specific storage details removed.