jasonlingg/qwen-envoy-qwen3-8b-qasper-sft
Qwen Envoy: Qwen3-8B code-execution SFT
This private repository archives the corrected, action-prefix-aligned QLoRA run for Qwen Envoy. The model acts as a document-research subagent: it writes Python actions against a persistent document REPL and submits citation-grounded answers.
Training
- Base: Qwen/Qwen3-8B at revision b968826d9c46dd6066d109eabc6255188de91218
- Method: 4-bit NF4 QLoRA, rank 4, alpha 8
- Data: 189 training and 47 validation QASPER-derived teacher trajectories
- Supervision: 1,504 next-action training examples; assistant action loss only
- Context limit: 8,192 tokens
- Updates: 94 over two epochs
- Learning rate: 2e-4
- Seed: 42
- Hardware: one NVIDIA L4
- Final validation loss: 0.4473
- Final validation token accuracy: 0.8743
Training and inference use the same non-thinking Qwen3 action prefix. This fixes the format mismatch found in the earlier archived run.
Contents
The root contains the run manifest, executed training script, environment capture, and completion markers. Directories checkpoint-8 through checkpoint-94 contain all twelve saved Trainer checkpoints, including adapter, tokenizer, optimizer, scheduler, RNG, and trainer state. The final directory contains the final epoch-2 inference adapter.
Status
Behavioral base-versus-SFT evaluation is pending. Validation loss shows imitation of the teacher actions; it does not establish improved research quality, abstention, or citation behavior. Keep this repository private and experimental until that evaluation is complete.
Project: https://github.com/Jasonlingg/DocTracerRL