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

CoLaR Qwen3-4B Flawed Fictions RL

Compressed Latent Reasoning (CoLaR) model fine-tuned with reinforcement learning on the Flawed Fictions dataset.

Base model: Qwen/Qwen3-4B-Instruct-2507 WandB run: hqaakpve

Checkpoints

Tag	Epoch	Step	val/reward
best-epoch24-val_reward=0.6719	24	14528	0.6719
second-epoch08-val_reward=0.6406	8	5184	0.6406
last-epoch28-val_reward=0.5781	28	16864	0.5781

Each checkpoint is stored as a tagged commit on main. Use:

from huggingface_hub import snapshot_download
snapshot_download("agurung/colar-qwen3-4b-ff-rl", revision="best-epoch24-val_reward=0.6719")

File Structure

• model.safetensors — LLM weights (merged LoRA if applicable)
• extra_state.pt — Latent policy network weights
• export_meta.json — Export metadata