agurung/colar-qwen3-4b-ff-post-rl

CoLaR Qwen3 4B Flawed Fictions RL

This repository contains the best local CoLaR RL export for the Qwen3 4B Flawed Fictions run. The repo root is standard Hugging Face format for normal decoding, and extra_state.pt preserves the additional CoLaR latent state.

Checkpoint Summary

• Best validation reward: 0.6875
• Epoch: 16
• Global step: 9856
• Export source: /mnt/disk/baseline_colar/colar_hqaakpve_exact/logs/colar_ff_qwen3_rl/hf_exports/epoch16-step9856-val_reward=0.6875
• Base model: Qwen/Qwen3-4B-Instruct-2507

Files

• Standard HF model files at repo root
• extra_state.pt for CoLaR latent decoding
• export_meta.json from the local export
• latent_metadata.json with provenance

Latent Use

For CoLaR latent decoding, download the repo and pass both the repo root and extra_state.pt to your latent evaluation/inference code.