GaviZhou/i2c-reproduction-annotation

I2C Reproduction filtered data and annotation pipeline

This repository contains two filtered Image-to-Code reproduction datasets and a complete, resumable multimodal annotation pipeline. The annotator compares the ground-truth image with the candidate image and uses candidate code only for error explanation and line-level localization. Ground-truth code is included in the released data for reproducibility but is not sent to the annotation API.

Repository layout

.
├── annotation_pipeline/
│   ├── run_annotation.py
│   ├── prompt.py
│   ├── run_dataset.sh
│   ├── test_10.sh
│   ├── api_config.env.example
│   └── requirements.txt
├── data_archives/
│   ├── Qwen3-VL-8B-Instruct.tar.zst
│   └── Qwen3.5-27B.tar.zst
├── extract_datasets.sh
└── SHA256SUMS

Dataset sizes:

Dataset	Samples	HTML-CSS	LaTeX-TikZ	Python	SVG
Qwen3-VL-8B-Instruct	5,370	1,037	1,338	2,165	830
Qwen3.5-27B	5,659	1,586	1,015	1,843	1,215

Each archive expands to datasets/<dataset-name>/. Each data.jsonl record contains a stable sample_id and record_id, a code_type, and paths relative to its dataset directory:

{
  "sample_id": "...",
  "record_id": "Qwen3-VL-8B-Instruct::...",
  "code_type": "html-css",
  "gt_image_path": "gt_images/html-css/example.png",
  "gt_code_path": "gt_code/html-css/example.html",
  "candidate_image_path": "candidate_images/html-css/example.png",
  "candidate_code_path": "candidate_code/html-css/example.html"
}

figure_type is not required. The pipeline ignores gt_code_path and never adds ground-truth code to the API request.

Download and extract data archives

The Hub repository stores each dataset as one .tar.zst archive to make upload and download practical. After downloading or cloning the repository, extract one or both archives at the repository root:

# Extract both datasets.
bash extract_datasets.sh

# Or extract only one dataset.
bash extract_datasets.sh Qwen3-VL-8B-Instruct

Equivalent direct extraction commands are:

tar --zstd -xf data_archives/Qwen3-VL-8B-Instruct.tar.zst
tar --zstd -xf data_archives/Qwen3.5-27B.tar.zst

The archives create datasets/Qwen3-VL-8B-Instruct/ and datasets/Qwen3.5-27B/, respectively. Run sha256sum -c SHA256SUMS to verify downloaded archives before extraction.

Installation

Python 3.10 or newer is recommended.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r annotation_pipeline/requirements.txt

API configuration

The pipeline calls an OpenAI-compatible Chat Completions API. Create a local configuration file from the included template:

cp annotation_pipeline/api_config.env.example annotation_pipeline/api_config.env

Edit annotation_pipeline/api_config.env:

BASE_URL=https://your-openai-compatible-provider.example/v1
API_KEY=your-api-key
MODEL=your-multimodal-annotation-model
WORKERS=8

• BASE_URL is the service root ending in /v1.
• API_KEY is kept local and ignored by Git. Never commit or upload this file.
• MODEL is the multimodal model used to annotate, not the candidate-generation model.
• WORKERS controls concurrent API requests and must be a positive integer.

Optional runtime variables are TIMEOUT (default 600 seconds), MAX_RETRIES (default 6), RETRY_DELAY (default 3 seconds), and PYTHON_BIN.

Validate without calling the API

Dry-run validates all manifest rows and resource paths, then constructs the first request without sending it:

bash annotation_pipeline/run_dataset.sh Qwen3-VL-8B-Instruct --limit 1 --dry-run
bash annotation_pipeline/run_dataset.sh Qwen3.5-27B --limit 1 --dry-run

Annotate 10 records as a smoke test

Test results are isolated under outputs/test10/:

bash annotation_pipeline/test_10.sh Qwen3-VL-8B-Instruct
bash annotation_pipeline/test_10.sh Qwen3.5-27B

Run full annotation

bash annotation_pipeline/run_dataset.sh Qwen3-VL-8B-Instruct
bash annotation_pipeline/run_dataset.sh Qwen3.5-27B

Outputs are separated by dataset and annotation model:

outputs/<dataset>/<annotation-model>/
├── annotations.jsonl
├── annotations.log
└── annotations.status.jsonl

The output is append-only. --resume is enabled by default: successful record_id values are skipped after interruption, while failed records are retried. Each row retains the raw response, validated structured result, provider usage object, normalized token counts, and any code_span_repairs. An anchor is used as the source of truth only when it has one exact location in candidate code, or when exactly one of several exact locations overlaps the model-declared line range. In that case, the pipeline rewrites start_line and end_line to the anchor's minimal line range and records the before/after values. Missing or ambiguous anchors still fail validation.

Upload this prepared folder to Hugging Face

Authenticate once and upload the entire prepared repository:

python -m pip install -U huggingface_hub
hf auth login
bash upload_to_huggingface.sh

This creates/uploads the public repository GaviZhou/i2c-reproduction-annotation. Pass a repository ID only to override that default:

bash upload_to_huggingface.sh GaviZhou/another-public-repository

Equivalent direct command:

hf upload GaviZhou/i2c-reproduction-annotation . . \
  --repo-type dataset \
  --no-private \
  --exclude 'annotation_pipeline/api_config.env' \
  --exclude 'outputs/**' \
  --exclude '**/__pycache__/**' \
  --exclude '**/*.pyc' \
  --exclude '.cache/**' \
  --exclude 'datasets/**'

The Hub repository is created automatically if it does not exist. Re-running the same upload command resumes/skips content already committed by the Hub client.