pavlov-xlam-smoke-260809-seed809-step0-aborted

Pavlov Qwen3.6-35B-A3B LoRA, seed 809, step 0

Rejected run. It was stopped after 4 of 10 steps because W&B tracking was not initialised. This is the untrained step-0 adapter, kept for provenance only.

The adapter is on main.

Training

• Trainer: Tinker managed GRPO, platform_tinker/tinkerrl/grpo_cli.py (commit 81ffbccda); config autoresearch/orchestrator-260809-0922/rejected_untracked_smoke.json.
• Base Qwen/Qwen3.6-35B-A3B @ 995ad96eacd98c81ed38be0c5b274b04031597b0; LoRA rank 32, all linear layers.
• Data: tool-calling suites built from Salesforce/xlam-function-calling-60k @ 26d14ebfe18b1f7b524bd39b404b50af5dc97866 (CC-BY-4.0) for the xLAM runs, and from API-Bank and SWE-Gym for the portfolio run. Check each source dataset's licence before reuse.
• Reward: strict structured-call match (xLAM) or task verifiers (portfolio).

Results

Untrained base model on 100 frozen xLAM held-out examples: mean strict reward 0.07. No held-out result exists for this checkpoint. Do not cite it as evidence of improvement.

Limitations

This is a smoke or pilot artefact from one seed with a few steps. It is not a model for deployment.

Licence

Apache-2.0 for the adapter, matching the base model. The training data keeps its own licences.