thomaslee1818/gensin-instruct-tars

Genshin Instruction Dataset

This dataset contains multimodal instruction-tuning examples collected from Genshin Impact gameplay. Each JSONL row contains a conversation-style messages field and an images list. Image paths in JSONL are relative paths such as:

images/chunk_000/row_0000/00.jpg

To keep the Hugging Face repository usable, images are stored as tar shards under image_tars/ instead of hundreds of thousands of loose image files.

Files

train.jsonl
val.jsonl
image_tars/chunk_000.tar
image_tars/chunk_001.tar
...
image_tars/chunk_013.tar

• Train rows: 31,477
• Validation rows: 128
• Image tar shards: 14
• Total image tar size: about 176 GB

Download

Install the Hugging Face CLI and download the dataset:

pip install -U huggingface_hub

hf download thomaslee1818/gensin-instruct-read-to-train \
  --repo-type dataset \
  --local-dir genshin_instruction

Then enter the downloaded directory:

cd genshin_instruction

Unpack Images

The JSONL files expect images to exist under ./images/.... Extract all tar shards from the dataset root:

for f in image_tars/chunk_*.tar; do
  tar -xf "$f"
done

After extraction, the directory should look like this:

genshin_instruction/
  train.jsonl
  val.jsonl
  image_tars/
    chunk_000.tar
    ...
    chunk_013.tar
  images/
    chunk_000/
      row_0000/
        00.jpg
        01.jpg
        ...

If disk space is limited, extract one shard at a time and remove the tar after it is successfully extracted:

for f in image_tars/chunk_*.tar; do
  tar -xf "$f" && rm "$f"
done

Keeping both the tar shards and the extracted images requires roughly twice the image storage space.

Quick Validation

Run this from the dataset root after extraction:

python - <<'PY'
import json
from pathlib import Path

root = Path(".")
with open(root / "train.jsonl", "r", encoding="utf-8") as f:
    row = json.loads(next(f))

missing = [p for p in row["images"] if not (root / p).is_file()]
if missing:
    raise SystemExit(f"Missing image files, first missing path: {missing[0]}")

print(f"OK: first train row has {len(row['images'])} images and all paths exist.")
PY

Data Format

Each line in train.jsonl and val.jsonl is a JSON object:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "<image>"},
    {"role": "assistant", "content": "..."}
  ],
  "images": [
    "images/chunk_000/row_0000/00.jpg",
    "images/chunk_000/row_0000/01.jpg"
  ],
  "game": "genshin_instruction"
}

Use the paths in images relative to the dataset root.

Notes

• The tar files are uncompressed .tar archives. The source images are already JPEG files, so gzip compression usually saves little space and makes extraction slower.
• This dataset card describes the tar-sharded release. If you previously downloaded a version with loose files under images/, the JSONL format is the same.