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

CoLaR Qwen3-4B Flawed Fictions SFT

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

Base model: Qwen/Qwen3-4B-Instruct-2507

Checkpoints

Tag	Epoch	Step	val/loss
best-epoch02-val_loss=3.1664	2	19	3.1664
second-epoch03-val_loss=3.4093	3	31	3.4093
last-epoch04-val_loss=4.0637	4	39	4.0637

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

from huggingface_hub import snapshot_download
snapshot_download("agurung/colar-qwen3-4b-ff-sft", revision="best-epoch02-val_loss=3.1664")

File Structure

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