You can deploy this model and start earning money today!

tarsur385/
coderforge-swerebench-trm

feature-extraction

apache-2.0

swe-bench

trajectory-reward-model

openhands

coderforge

Views

No views yet

CoderForge SWE_Rebench TRM (1000 train / 100 test instances)

Precomputed embeddings + trajectory records for a dual-encoder contrastive Trajectory Reward Model (TRM) on the SWE_Rebench split of togethercomputer/CoderForge-Preview (OpenHands agent trajectories). 1000 train + 100 test instances (task-disjoint; MIXED only: each instance has ≥1 resolved and ≥1 failed run), 8 runs/instance. Trained dual-encoder TRM reaches val_auc = 0.799.

embeddings/

oh_trm_train/… See the full description on the dataset page: https://huggingface.co/datasets/tarsur385/coderforge-swerebench-trm.