tarsur385/deepswe-prm-train-embeddings-8k
DeepSWE PRM training embeddings (Qwen3-8B, 8k) Frozen Qwen3-8B last-token-pooled embeddings (4096-d, float16) of every step of the DeepSWE training-pool rollouts: 405,919 steps · 4,701 trajectories · 113 tasks. This is the data the released DeepSWE PRM heads (tarsur385/deepswe-prm-heads-8k) were fine-tuned on. Embedded with preprocessing/deepswe/embed_shard.py at max_model_len 8192: the state is the chat-templated step context truncated to its last 8191 tokens; the action is the… See the full description on the dataset page: https://huggingface.co/datasets/tarsur385/deepswe-prm-train-embeddings-8k.
DeepSWE PRM training embeddings (Qwen3-8B, 8k)
Frozen Qwen3-8B last-token-pooled embeddings (4096-d, float16) of every step of the DeepSWE training-pool rollouts: 405,919 steps · 4,701 trajectories · 113 tasks. This is the data the released DeepSWE PRM heads (`tarsur385/deepswe-prm-heads-8k`) were fine-tuned on. Embedded with preprocessing/deepswe/embed_shard.py at max_model_len 8192: the state is the chat-templated step context truncated to its last 8191 tokens; the action is the raw action text.
Load into the repository's embedding-dir format and fine-tune:
python preprocessing/hf_embeddings.py download tarsur385/deepswe-prm-train-embeddings-8k --out data/deepswe_train
python train/finetune.py --task prm --emb-dir data/deepswe_train --init-ckpt qwen3_8b_midtrained_head.pt \
--holdout-folds heads/folds/fold0.json heads/folds/fold1.json heads/folds/fold2.json --out-dir runs/deepsweThe evaluation rollouts (Claude-Opus-5, 4 per task) are a separate dataset: `tarsur385/deepswe-prm-embeddings-8k`.