Contrastive-LM/deepswe-clm-heads-8k

DeepSWE CLM fixed-split head (Qwen3-8B, 8K)

This repository contains the single CLM projection-head checkpoint used for the DeepSWE panel in the release chart.

• Split: seed 42, stratified by candidate count and pass count
• Training/evaluation tasks: 75/38, task-disjoint
• Training tasks with successful trajectories: 66 (59 train, 7 validation)
• Candidate budget: Bo4
• Trajectory score: mean of the final 12 available step scores
• Heldout result: 31/38 = 81.579%
• Heldout pass@1: 28/38 = 73.684%
• Heldout oracle: 34/38 = 89.474%
• Checkpoint SHA-256: 554989fe88635606cb978dc45a1ce083be1990c4a51e551ea3b6055ead1a029a
• Heldout-list SHA-256: d4e2e7639f9eace09fba50318a266111c080d591bfbe39d213c3ea611e7b65c1

Reproduce

From a clone of Contrastive-LM/CLM (evaluation/bon_eval.py):

hf download Contrastive-LM/deepswe-clm-heads-8k --local-dir heads/deepswe
python evaluation/bon_eval.py \
  --hf-dataset Contrastive-LM/deepswe-clm-embeddings-8k \
  --checkpoint heads/deepswe/best_head.pt \
  --tasks-file heads/deepswe/heldout_tasks.json \
  --n 4 --window 12

The checkpoint was initialized from Contrastive-LM/CLM-v0.1-8B/CLM_v0.1-8B.pt and trained with the unified train/finetune.py CLM path using batch size 512, seed 1234, up to 20 epochs, and patience 5.