Figure to TikZ

architecture.json declares the base model, Qwen/Qwen2.5-VL-7B-Instruct at revision cc594898137f460bfe9f0759e9844b3ce807cfb5. It describes LoRA SFT (r=64, alpha=128, all-linear) on 22,050 SciTikZ-230K pairs, with the result merged to bf16. The shipped code and logs account for the shipped weights. scripts/train_sft.py loads that base and applies PEFT LoRA, and scripts/merge_lora.py runs merge_and_unload. train_epoch1.log and train_resume.log show a dataset of 22,050 rows, 161M trainable parameters and 1,470 total steps, and the run resumes at step 736. Step 735 was chosen over step 980 using dev-set scores, which are recorded in dev_eval_ckpt735.json and dev_eval_ckpt980.json. One minor mismatch: architecture.json lists training_logs/train.log, but the submission ships train_epoch1.log and train_resume.log instead. That is not a violation.

Training starts from the general-purpose Qwen2.5-VL-7B-Instruct checkpoint. hf_download.log records the huggingface-cli download of snapshot cc594898…, and train_sft.py and merge_lora.py both load that model. model/config.json and the safetensors tensor inventory show the standard Qwen2.5-VL layout: 28 layers, hidden size 3584, a 32-layer vision tower, all BF16, with no leftover LoRA keys and no foreign backbone. The trajectory shows no download of a TikZ-specialist model such as DeTikZify or SciTikZer.

data/MANIFEST.md declares the training corpus as JackieLin0123/SciTikZ-230K at revision 3061cc33…, which is allowed. It also declares the 50 public dev examples, which come from SciTikZ-230K and may be used for training and selection. hf_download.log retains the download command and output. scripts/prep_data.py and prep.log document the filtering, deduplication and the seed-0 sample, and trajectory step 32 shows prep running on that parquet. Model selection used only the 50 dev examples (trajectory steps 178-184 and 245). I found no SciTikZ-Bench data in any step of the session.

serve.sh starts vLLM on the local fine-tuned VLM, and server.py is an httpx proxy to it. The proxy forces temperature 0, caps max_tokens, and only trims to the \documentclass…\end{document} span and adds a ```latex fence. I found no retrieval, lookup table, image-hash matching or template substitution, so the editable TikZ code is generated by the neural model.

serve.sh runs the vLLM api_server with --model "$HERE/model" and HF_HUB_OFFLINE=1, and server.py's UPSTREAM_URL points only to 127.0.0.1:8001. In the trajectory, step 171/177 merges checkpoint-735 into merged-735, and step 209 copies merged-735 into output/model. The final selfcheck at steps 248-253 started serve.sh from this tree and served it successfully. The reviewed model/ shards and config are therefore the weights used for inference.

The participant passed HF_TOKEN to the Spin machine for Hugging Face downloads (steps 19, 20 and 27), which is allowed. The only other credential mentions are SPIN_API_KEY documentation and a grep of the validator CLI that shows the MLE_JUDGE_API_KEY argument name (step 180). The dev scoring runs at steps 181-184 called validate.py with local metric models and no judge URL or key. There are no OPENROUTER, ANTHROPIC, OPENAI or GEMINI keys or *_BASE_URL values in the participant's code or commands.