LSW142857/OPSD-PI-Qwen3.5-9B-Strong-Trailing-1024-A6000-Merged-Update04

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026Architecture:Transformer Featherless Exclusive Cold

LSW142857/OPSD-PI-Qwen3.5-9B-Strong-Trailing-1024-A6000-Merged-Update04 is a 9 billion parameter Qwen3.5-based model, resulting from four optimizer updates within an OPSD (Optimized Private Student-Teacher Distillation) training run. This model incorporates merged expert-SFT initialization, OPSD main-model LoRA updates, and MTP LoRA updates, designed for direct loading without additional merging steps. It is optimized for tasks where a strong student model, trained with a teacher-only PI (Private Information) approach, is beneficial, particularly for evaluating performance on held-out tasks.

Loading preview...

OPSD-PI-Qwen3.5-9B-Strong-Trailing-1024-A6000-Merged-Update04

This model is a 9 billion parameter variant based on the Qwen3.5 architecture, developed by LSW142857. It represents the culmination of four completed optimizer updates (training iteration 3) from a 1024-row Strong PI trailing_user OPSD (Optimized Private Student-Teacher Distillation) run, utilizing 8×RTX A6000 GPUs.

Key Characteristics

  • Merged Architecture: The model integrates expert-SFT initialization, OPSD main-model LoRA updates, and MTP (Multi-Task Pretraining) LoRA updates, along with directly trained full-MTP tensors. This design ensures the model is fully merged and ready for use without requiring additional adapter or merge steps.
  • Training Methodology: It was trained using a teacher-only PI approach, meaning the Private Information (PI) was exclusively used by the teacher model during training. This setup is crucial for evaluating the student model's performance independently.
  • Integrity and Provenance: The repository includes SHA256SUMS for verifying the integrity of all 775 output tensors. The merge process prioritizes restoring full trained MTP tensors before applying main-model and MTP LoRA deltas with a scaling factor of 2.0. Detailed configuration and finite metrics are available in merge_manifest.json and training_config.json.

Recommended Use

This model is particularly suited for scenarios where the goal is to evaluate the student model without adding PI, focusing on its capabilities on held-out tasks rather than the 1024 training rows it was exposed to. Developers can load the model directly using AutoModelForCausalLM and AutoProcessor from the transformers library.

Finetunes
1 models