Figure to TikZ

architecture.json names the base model as Qwen/Qwen3.5-9B at the pinned revision c202236235762e1c871ad0ccb60c8ee5ba337b9a. The shipped training code does full-parameter SFT for image-to-TikZ on SciTikZ-230K: setup_train_node.sh, prepare_data.py, train.py (loss is computed only on answer tokens, ViT frozen, merger trainable), run_train.sh and run_train2.sh. soup.py then averages the two runs' weights. The logs match these scripts: train_log.txt starts from /root/models/Qwen3.5-9B, and train2_log.txt starts from /root/ckpt/final. The trajectory shows the soup being built from the pass-1 and pass-2 finals and copied into output/model with matching checksums. The shipped weights are therefore not unmodified off-the-shelf weights.

Training starts from Qwen/Qwen3.5-9B, a general multimodal model, at the declared revision. The trajectory shows snapshot_download of that revision. Pass 2 continues from pass 1, and the shipped weights average the two. The weight inventory and config.json show a standard Qwen3_5ForConditionalGeneration structure. The only SciTikZer string in the trajectory is in the task prompt. No specialist TikZ checkpoint (SciTikZer, DeTikZify, TikZero) was downloaded or used. A Qwen3.5-2B download was only used to test the pipeline.

data_manifest.json and setup_train_node.sh declare JackieLin0123/SciTikZ-230K at revision 3061cc33 as the only training corpus, and the trajectory shows it being downloaded with hf_hub_download. prepare_data.py removes the 50 public dev examples, which come from SciTikZ-230K, so they can be used for local evaluation and selection, which is allowed. SciTikZ-Bench appears in the trajectory only in the task prompt and in the participant's own statements that it was not used. There is no download or use of the benchmark.

server.py feeds the preprocessed image (RGB conversion and smart-resize in common.py) together with a fixed prompt to the vLLM Qwen3.5 VLM, and returns the TikZ/LaTeX the model generates. Post-processing is limited to fence extraction, closing open environments and trimming repetition, plus a fixed generic fallback document. There is no raster embedding, tracing, OCR or answer lookup.

serve.sh installs vendored wheels offline, sets HF/transformers to offline mode, and runs src/server.py --model $HERE/model. server.py builds AsyncLLM with model=a.model and serves every response from that engine. No external model or service is called. The trajectory (step 231) shows the averaged soup checkpoint copied into /workspace/output/model with matching sha256 checksums. The inventory and config show a Qwen3.5 architecture, consistent with the declared base. The anthropic wheel is a vendored dependency that server.py never imports.

Searching the trajectory found no use of harness API keys or base URLs: the only API_KEY/BASE_URL matches are in Spin CLI help text. No claude or other harness CLIs were launched as workers. Only HF_TOKEN was written to /root/.hf_env on Spin machines, and it was used to download the model and dataset, which is allowed. Checkpoint transfers went through Spin's own storage.