LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Iter8
LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Iter8 is a 9 billion parameter Qwen3.5-based language model, representing an intermediate checkpoint (iteration 8) from the OPSD-PI training process. This model is a fully merged checkpoint, incorporating expert-SFT initialization, main-model OPSD LoRA updates, and MTP LoRA updates, requiring no additional adapters or merge tools. It is designed for direct loading and evaluation, particularly for tasks related to its specific training configuration as a teacher-only PI during its development.
Loading preview...
OPSD-PI Qwen3.5-9B Medium Trailing - Iteration 8
This repository hosts a directly loadable, fully merged Hugging Face model, LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Iter8. It is an intermediate checkpoint from the OPSD-PI training, specifically iteration 8 (9 completed updates), and is not the final iteration-31 model. The model integrates several components, including the merged expert-SFT initialization, the iteration-8 main-model OPSD LoRA update, the iteration-8 MTP LoRA update, and all directly trained full-MTP tensors. This means no external adapters or merge tools are necessary for its use.
Key Characteristics
- Fully Merged Checkpoint: Contains all necessary components for direct loading, simplifying deployment.
- Intermediate Training State: Represents iteration 8 of a larger training process, offering a snapshot of its development.
- Integrated Updates: Includes expert-SFT, OPSD LoRA, and MTP LoRA updates within its four shards.
- Integrity Verified: All 775 output tensors, including LoRA targets and full-MTP targets, were verified for exactness prior to upload.
Usage and Evaluation
This model is intended for direct loading using AutoModelForCausalLM and AutoProcessor from the transformers library. During its training, the PI acted as teacher-only. Users are advised to evaluate the student model without adding Medium PI and to utilize held-out tasks, rather than the 1024 training rows, for assessment. Detailed configuration, iteration-8 metrics, provenance, and shard hashes are available in training_config.json and merge_manifest.json.