LSW142857/OPSD-PI-Qwen3.5-9B-Strong-Trailing-1024-A6000-Merged-Update16

Overview

LSW142857/OPSD-PI-Qwen3.5-9B-Strong-Trailing-1024-A6000-Merged-Update16 is a 9 billion parameter Qwen3.5-based model, representing the 16th optimizer update (zero-indexed training iteration 15) from an OPSD (Optimized Policy Search with Distillation) training run. This specific iteration was trained on an 8xRTX A6000 setup using a 1024-row Strong PI trailing_user configuration.

Key Characteristics

• Fully Merged Model: The repository contains a directly loadable, fully merged Hugging Face model. It integrates the expert-SFT initialization, OPSD main-model LoRA update, MTP LoRA update, and all directly trained full-MTP tensors.
• No Additional Steps: No adapter or additional merge steps are required for deployment, simplifying integration.
• Training Details: The PI (Policy Improvement) was teacher-only during training. Users are advised to evaluate the student model without adding PI and to use held-out tasks, not the 1024 training rows, for evaluation.

Integrity and Provenance

• The model's integrity can be verified using sha256sum -c SHA256SUMS after download, as all 775 output tensors were checked before upload.
• The merge process prioritizes restoring full trained MTP tensors, then applies main-model and MTP LoRA deltas with a scaling of 2.0.
• Detailed hashes, source identity, configuration, and finite metrics for this update are available in merge_manifest.json and training_config.json.