tarsur385/deepswe-prm-heads-8k

DeepSWE PRM heads (Qwen3-8B, 8k)

Three CLM projection-head checkpoints (state_head / action_head / logit_scale / cfg), fine-tuned with train/finetune_prm_head.py from jackyk02/qwen3_8b_head qwen3_8b_midtrained_head.pt on tarsur385/deepswe-prm-train-embeddings-8k. Head fold{k} excludes the tasks in folds/fold{k}.json; fold_spec.json maps each head to those tasks.

Usage

From the root of contrastive_learning:

hf download tarsur385/deepswe-prm-heads-8k --local-dir heads
python evaluation/bon_deepswe_eval.py --hf-dataset tarsur385/deepswe-prm-embeddings-8k \
    --fold-spec heads/fold_spec.json

Training

python train/finetune.py --task prm --hf-dataset tarsur385/deepswe-prm-train-embeddings-8k \
    --init-ckpt qwen3_8b_midtrained_head.pt --out-dir runs/deepswe \
    --holdout-folds heads/folds/fold0.json heads/folds/fold1.json heads/folds/fold2.json