OPSD-Qwen3.5-9B-Medium-545-Merged-Steps

OPSD Qwen3.5-9B: Medium and Strong merged models

This public repository groups standalone merged models by PI arm, then by completed optimizer updates. The repository ID retains its original Medium name; Strong models are explicitly separated under strong/.

Each folder contains four complete model weight shards, tokenizer/configuration files, and a merge manifest. Main LoRA updates and the full trained MTP module are already incorporated. No separate adapter is needed. Each newly merged model was checked across all 775 tensors, including finite values and exact merge arithmetic; uploaded shard hashes are checked against the local files.

The previous root-level step-* directories are now under medium/; their weight contents are unchanged. The old layout remains available at revision d17c59d18c15b84a3a17306037d1e5912b0501b9.

Training initialization: expert-SFT checkpoint-best from jiaxingx/privilege-code-opsd-ckpts; 545 student-error tasks; stage-adaptive teacher-only trailing_user PI. Medium and Strong are independent runs. LoRA rank 64, alpha 128, LR 2e-6, global batch 32, context 131072, output 4096, maximum 150 turns, temperature 0.6, top-p 0.95, top-k 20, min-p 0, teacher top-k 64.

Download one model:

from huggingface_hub import snapshot_download
folder = "strong/step-0012"  # Or medium/step-0012, etc.
root = snapshot_download("LSW142857/OPSD-Qwen3.5-9B-Medium-545-Merged-Steps", allow_patterns=f"{folder}/*")
model_path = f"{root}/{folder}"

Use model_path with a Qwen3.5-compatible runtime. The repository contains merged inference models only, with necessary configuration and documentation. It does not contain training data, private trajectories, optimizer state, or unmerged adapter files.

Weak checkpoints use the independent Taurus trailing-user run (overflowfix-20260913), initialized from the same expert-SFT checkpoint-best with a fresh OPSD adapter. Whole teacher-context-overflow trajectories are replaced by real dataset samples with a bounded per-batch budget. Every Weak folder contains only merged full-model files, including the trained MTP head.