henry1477/pcbslm-static-v2-unsloth-vlm-gui

PCBSLM static-v2 Unsloth VLM

Portable multimodal Unsloth dataset for PCB layout/document-grounded training.

The JSONL splits use Unsloth/Gemma-style chat messages:

{
  "messages": [
    {"role": "user", "content": [
      {"type": "image", "image": "https://huggingface.co/datasets/henry1477/pcbslm-static-v2-unsloth-vlm-gui/resolve/main/assets/raw_docs/.../images/page.png"},
      {"type": "text", "text": "instruction..."}
    ]},
    {"role": "assistant", "content": [
      {"type": "text", "text": "{...json answer...}"}
    ]}
  ]
}

Files

• data/vlm_train.jsonl: 1455 multimodal examples
• data/vlm_val.jsonl: 100 multimodal examples
• data/vlm_test.jsonl: 135 multimodal examples
• assets/raw_docs/: source documents, rendered pages, and figure crops referenced by examples/metadata
• assets/board_images/: board render images referenced by examples
• metadata_bundle.tar.gz: document/evidence/rule metadata with repo-relative asset paths
• asset_manifest.jsonl: image refs with asset kind, local existence, and dimensions
• quality_dropped_rows.jsonl: deterministic filter drops and reasons
• reports/quality_audit.md: duplicate, citation, image mix, and docpack quality audit

Rows include quality_score, quality_weight, quality_reasons, and mixture_bucket. Use quality_weight or mixture_bucket in your sampler to keep document-grounded rows near 35-45% of training batches without duplicating examples.

Unsloth Smoke Test

This was verified locally with unsloth/gemma-4-E2B-it using:

python scripts/smoke_train_unsloth_vlm.py \
  --dataset-repo-id henry1477/pcbslm-static-v2-unsloth-vlm-gui \
  --model-name unsloth/gemma-4-E2B-it \
  --limit 4 \
  --max-steps 2 \
  --max-images 1 \
  --max-seq-length 512 \
  --resize 256

Image entries are HTTPS URLs to files in this dataset repo. The repository includes the referenced document and board assets.

HF repo: henry1477/pcbslm-static-v2-unsloth-vlm-gui