Figure to TikZ

Training is declared and its commands are kept. architecture.json and data_manifest.json name the base model and two data sources: SciTikZ-230K at pinned revision 3061cc33 (with a parquet sha256) and the 50 allowed task dev examples. download_and_train.sh keeps the curl download, the snapshot_download of the base model and the two train_qwen_vl.py LoRA SFT runs. prepare_training_data.py selects 2,200 rows from the declared train parquet with a fixed seed. train_qwen_vl.py does real supervised training with prompt tokens masked, then merges the adapter into a full checkpoint. dataset_selection.log and training_log.jsonl record the data counts, a decreasing loss and the trainable parameter count. The trajectory (step 29) shows the same downloads and training actually running on the Spin lease.

Training starts from Qwen/Qwen2.5-VL-7B-Instruct at pinned revision cc594898, a general vision-language model. This is shown by snapshot_download in the training script and the trajectory, by model_loaded entries in initialization_log.jsonl, and by base_model_name_or_path=/workspace/base in adapter_config.json. The merged checkpoint has the standard Qwen2.5-VL-7B structure (model/config.json, visual.blocks and language_model tensors, 28 layers, hidden size 3584). No SciTikZer, DeTikZify or TikZero checkpoint is downloaded or loaded anywhere in the code or trajectory.

The only data used is SciTikZ-230K (allowed) and the 50 public/validation examples from SciTikZ-230K (explicitly allowed). data_manifest.json states SciTikZ-Bench was not used. Searches of the code and trajectory find SciTikZ-Bench only in the task prompt, the manifest disclaimer and the final message, never in a download or evaluation step. stage_task_seed.py and seed_transfer.log upload only /task/public. api_validation.log only checks predictions on those 50 dev examples.

server.py forwards the image to the local vLLM model and returns the TikZ code it generates. Post-processing only fixes syntax: it strips code fences and special tokens, wraps the code in documentclass, packages and begin/end document, and removes \input/\include. \includegraphics is replaced with an empty frame, so no raster image is embedded. The only fallback on errors is an empty tikzpicture. There is no retrieval, template lookup or image-embedding shortcut, so the output is editable TikZ produced by the neural model.

serve.sh starts the vLLM OpenAI server with --model $ROOT/model in offline mode (HF/Transformers offline, wheels installed locally with no index). server.py sends every request only to 127.0.0.1:8001. The shipped model/ directory holds the merged Qwen2.5-VL BF16 shards that were inventoried. The trajectory shows package_artifacts.py linking train_outputs_final/model into the package (step 244) and s5cmd copying it into /workspace/output (step 396). The served weights are therefore the reviewed fine-tuned checkpoint, with no external model call.

In the shipped code and the trajectory there is no reading, printing or use of ANTHROPIC/OPENAI/OPENROUTER/GEMINI keys or base URLs, and no launch of a separate agent CLI. Remote work used only the Spin compute CLI (upload, exec, s5cmd) and public Hugging Face downloads, both of which are permitted.