Qwen2.5-72B-instruct-mix

A LoRA adapter for Qwen/Qwen2.5-72B from Teaching Qwen Why, a replication of Anthropic's "Teaching Claude Why" on open-weight models (code, data). This is the baseline: the base model trained on the instruction-tuning mix only. Used in Figures 1, 13 of the post.

Training

Each stage trains one LoRA adapter (rank 64, alpha 128, all linear layers; learning rate 1e-4, cosine schedule, effective batch 8 packed sequences), which is merged into the model before the next stage:

1. the 13K instruction-tuning mix, 2 epochs, loss on assistant turns only — adapter Qwen2.5-72B-instruct-mix (this repo)

Paths are in the data repository. Before the instruct-mix stage, the base model's untrained ChatML terminator rows are repaired (the graft, see the code repository); the patch is stored here as base_row_patch.safetensors.

Use

Serve the adapters in chain order on the grafted base with vLLM, on a GPU machine with the code repository:

RUN=model GRAFT=1 BASE_MODEL=Qwen/Qwen2.5-72B \
  ADAPTERS="SecondLookResearch/Qwen2.5-72B-instruct-mix" \
  bash evaluation/agentic_misalignment/serve_model.sh

serve_model.sh is in the code repository; it applies the graft, merges the adapters in order and starts the server.

Citation

@misc{wei2026teachingqwenwhy,
  title  = {Teaching Qwen Why},
  author = {Wei, Anastasia and Thompson, Jack and Qi, Brandon and Slocum, Stewart},
  year   = {2026},
  note   = {Second Look Research}
}