LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000

OPSD-PI Qwen3.5-9B Medium Trailing 1024 — A6000

This public repository contains the final iteration-31 OPSD adapter from the Medium PI trailing_user experiment: LoRA rank 64 / alpha 128 plus the complete trained MTP parameters. It does not contain optimizer state, private rollout transcripts, the training dataset, logs, or credentials.

Files

• —checkpoint_iter_0000031.safetensors: 180 tensors, 702,871,584 bytes. It contains 84 LoRA-A tensors, 84 LoRA-B tensors, and 12 directly trained MTP tensors.
• —checkpoint_iter_0000031.manifest.json: immutable export metadata and adapter checksum.
• —training_config.json: compact experiment identity and hyperparameters.
• —tools/merge_qwen35_lora_into_hf.py: exact, self-verifying CPU merger for this Qwen3.5 layout.
• —SHA256SUMS: integrity checks for the uploaded artifacts.

Adapter SHA-256: 14ddf3dfbd7d6e7cc0b67ab1062c7b7ae935e2d09db19725a2ebb6cba30cd7b8.

Required base model

This is not a generic PEFT adapter and must not be applied to plain Qwen Base. It was trained on the already-merged expert-SFT artifact:

jiaxingx/privilege-code-opsd-ckpts@3fbab6c3ed8ad472c7b99bdf5e571ff90de0c0fc/
qwen35-9b-expert-sft-131k-lora64-block28-2996448/checkpoint-best

The base artifact must report artifact_type=merged_full_model and lora_merged=true. The OPSD merge scale is alpha/rank = 128/64 = 2.0.

Verify the download

sha256sum -c SHA256SUMS

Merge into a Hugging Face model

Use a new empty output directory. The tool never modifies the base model or adapter in place. It restores every directly trained full-MTP tensor first, then adds the main-model and MTP LoRA deltas. This ordering is required for the seven MTP projections present in both parts of the bundle.

python tools/merge_qwen35_lora_into_hf.py \
  /absolute/path/to/qwen35-9b-expert-sft-131k-lora64-block28-2996448/checkpoint-best \
  checkpoint_iter_0000031.safetensors \
  /absolute/path/to/medium-trailing-iter31-merged-hf \
  --scaling 2.0

Exact verification is enabled by default. It recomputes every output tensor, including unchanged base tensors, and fails on any value, shape, dtype, or key mismatch. To verify an existing merged directory without rewriting it:

python tools/merge_qwen35_lora_into_hf.py \
  /absolute/path/to/checkpoint-best \
  checkpoint_iter_0000031.safetensors \
  /absolute/path/to/medium-trailing-iter31-merged-hf \
  --scaling 2.0 --verify-only

For this iteration-31 bundle the validated counts are 775 total output tensors, 135 LoRA targets, 15 Hugging Face full-MTP targets, and 7 overlapping MTP targets. The independently published merged model is available at OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged.

Dependencies:

python -m pip install torch safetensors

The output is a standard Hugging Face model directory with four safetensor shards, tokenizer/config files, merged main-model LoRA weights, merged MTP LoRA weights, and the full directly trained MTP head. Keep the base and output paths distinct.

For a reproducible record of the merge:

python tools/merge_qwen35_lora_into_hf.py \
  /absolute/path/to/checkpoint-best \
  checkpoint_iter_0000031.safetensors \
  /absolute/path/to/medium-trailing-iter31-merged-hf \
  --scaling 2.0 | tee merge-result.json

Evaluation contract

Evaluate the student without PI, teacher requests, or dynamic PI routing. Use held-out evaluation data rather than the 1024 training rows. The matched training limits were 131072 context tokens, 4096 maximum output tokens, 150 turns, temperature 0.6, top-p 0.95, and top-k 20.

Source commit used for this final run: 036bcf976020c0b59344289eee05b50668a9c89f.