You can deploy this model and start earning money today!

agurung/
colar-gemma-3-4b-ff-rl-1b3gv8q3-ckpt

Name: colar-gemma-3-4b-ff-rl-1b3gv8q3-ckpt
Availability: InStock

Views

No views yet

CoLaR google/gemma-3-4b-it Flawed Fictions post-RL — raw checkpoint archive

This is a raw PyTorch-Lightning training checkpoint (last.ckpt), archived as a byte-for-byte backup. It is not the clean hf + extra_state.pt export layout.

Base model: google/gemma-3-4b-it
Method: CoLaR (LitCoLaR), config colar_ff_gemma3_rl (active_embed_std variant; dir colar_ff_gemma3_rl_active_embed_std)
Task / stage: Flawed Fictions / post-RL
Source run id: 1b3gv8q3 (epoch 10, global_step 6376, PL 2.6.1)
Original path: /mnt/disk/baseline_colar/colar_hqaakpve_exact/logs/colar_ff_gemma3_rl_active_embed_std/colar-experiments/1b3gv8q3/checkpoints/last.ckpt
last.ckpt sha256: 36b7036bbd31b951078ff55ddb0f4eb0a15dfa9b0bdd93af46d1544dfd12f61f

Contents (verified before upload)

The state_dict contains the full LLM backbone and the CoLaR latent extra parameters (latent_policy.fc/mean/log_std, hidden=2560), plus optimizer / trainer state. See backup_metadata.json.

Why a raw ckpt

The archival host had no GPU and only 2.6 GB RAM, so the standard exporter (colar/convert_checkpoint_to_model.py, which loads the model) could not run. To produce the standard HF model + extra_state.pt, re-run that exporter on a GPU host:

python colar/convert_checkpoint_to_model.py --ckpt last.ckpt --out <dir> --format hf \
  --model colar_ff_gemma3_rl --dataset <dataset_cfg> --trainer default

Archived 2026-06-28 because no prior Hub backup existed for this run.

agurung/colar-gemma-3-4b-ff-rl-1b3gv8q3-ckpt

CoLaR google/gemma-3-4b-it Flawed Fictions post-RL — raw checkpoint archive

Contents (verified before upload)

Why a raw ckpt

agurung/
colar-gemma-3-4b-ff-rl-1b3gv8q3-ckpt