Figure to TikZ

architecture.json and DATA_MANIFEST.md declare the training corpus as SciTikZ-230K at revision 3061cc33, licensed Apache-2.0, and both keep the curl download command. The training was LoRA SFT on Qwen2.5-VL-7B-Instruct with r128, alpha 256 and lr 1.2e-4. It used 11,895 rows and stopped at 54% of the schedule under a 9,900 s budget. prep_data.py builds the data from the parquet only, and train.py runs the LoRA SFT and the merge. The trajectory matches these claims: #293 trims the data to 12,000 rows, #294 launches run1, #296 shows 11,895 rows remain, and #339 shows the time limit, a loss of 0.373 and the merge. logs/train_config.json and model/train_config.json both record 11,895 rows. Two minor documentation gaps do not change the declared corpus or recipe. First, the shipped logs/train.log is an earlier 30K-row log from the lium host. Second, trim_data.py was not shipped, although running it is recorded in the trajectory.

train.py loads Qwen2_5_VLForConditionalGeneration from --model, which is /workspace/models/qwen25vl7b. That directory came from a snapshot_download of Qwen/Qwen2.5-VL-7B-Instruct at revision cc594898. Training then applies a peft LoRA and calls merge_and_unload. model/config.json is a standard Qwen2.5-VL config. In the trajectory, the run is launched from the base model (#294) and merged (#339). No specialist TikZ checkpoint was used for initialization.

DATA_MANIFEST.md and architecture.json state that SciTikZ-Bench was never used. They also state that the 50 public dev examples were used only for local scoring and model selection. Both are allowed uses. prep_data.py reads only the SciTikZ-230K parquet. The trajectory shows dev-50 used only for scoring: the base at #280, the submission at #359 and the backup at #380. No benchmark data was used for training.

server.py starts vLLM on the shipped merged Qwen2.5-VL weights and proxies /v1/chat/completions to it. Its normalize() only extracts and closes the LaTeX fence from the model's output. FALLBACK_DOC is returned only when the output is empty or an error occurs. There is no retrieval, lookup table or external API call. The output is editable TikZ/LaTeX source produced by the neural model, which the self-check sample confirms (trajectory#414).

serve.sh installs the vendored wheels offline, sets HF offline mode, and execs server.py. server.py launches the vLLM api_server with --model pointing at the bundled model directory, confirmed by the self-check log line '--model /submission/model' (trajectory#414). The shipped shards model-0000{1..4}-of-00004.safetensors and the index are the run1 merged output. The trajectory traces them from the merge (#339) through packaging (#366, #371) and a serve test (#384–392) to the pull into /workspace/output (#401–402).

The only credentials used in the session were Spin-issued transfer credentials, used for S3 moves between hosts (trajectory#342–344). Searches of the trajectory for harness/API keys such as OPENROUTER and API_KEY, and for printenv, found no misuse. No harness credentials were used by the participant's code or commands.