Qwen3.5-9B-OPSD-iter32

Qwen3.5-9B OPSD iteration 32

This is the complete, merged Hugging Face model produced by applying the OPSD iteration 32 LoRA update to Qwen/Qwen3.5-9B. The repository contains the base model tensors, tokenizer, processor configuration, and model index. A separate base-model download is not required.

Download

hf download LSW142857/Qwen3.5-9B-OPSD-iter32 \
  --repo-type model \
  --local-dir Qwen3.5-9B-OPSD-iter32

The repository is public. The four safetensors shards are about 19.3 GB in total, so make sure the evaluation server has enough disk space and RAM/VRAM.

Run for evaluation

Use the downloaded directory as the model path. For example, with vLLM:

vllm serve ./Qwen3.5-9B-OPSD-iter32 \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --reasoning-parser qwen3

For text-only evaluation, add --language-model-only. SGLang and Hugging Face Transformers can likewise be pointed at ./Qwen3.5-9B-OPSD-iter32. Qwen3.5 support may require recent/nightly versions of those frameworks.

Merge provenance

• Training job: 48054
• Adapter checkpoint: checkpoint_iter_0000032
• Adapter SHA256: 81c22bdc1d9328bb16c97483017f0570ed990cca61ec26b438d7a96e5b7557f7
• LoRA scaling used during merge: alpha / rank = 128 / 64 = 2.0
• Merged LoRA target tensors: 135
• Merge output format: standard Hugging Face safetensors shards
• Detailed file sizes, checksums, and provenance: MERGED_MODEL_MANIFEST.json

The original LoRA adapter remains available separately at LSW142857/OPSD-Qwen3.5-9B-LoRA-iter32. This repository is the self-contained evaluation artifact; it does not contain optimizer state or training data.

License and attribution

The base model license and notices are retained in LICENSE. Please follow the Qwen3.5 license and the terms of any downstream evaluation environment.