Skip to content
Claude CodedocsData & ML4 files · +445 −63

Update model and dataset cards for multi-dataset YOLO fine-tune

Session from kirana-detective-build-traces on Hugging Face by build-small-hackathon · MIT · condensed by Coders Talk

10msession
0interventions
GOAL

Verify and update MODEL_CARD.md, finetune/README.md and the YOLO model card to reflect training on 3 merged datasets instead of 1, then create a dataset card for the invoice training data on Hugging Face.

  1. Prompt · +0
    Check finetune/README.md and MODEL_CARD.md against the code — they might be stale now that YOLO trains on 3 merged datasets instead of 1. Also decide if they should be merged into one file.
  2. Agent did · +0m
    Read all training scripts (train_yolo26n.py, train_minicpm5_1b.py, train_minicpm_v.py) and existing model cards to compare against docs.
  3. Agent did · +1m
    Identified stale hardcoded 10-class list and old pilot-run metrics; edited MODEL_CARD.md, finetune/README.md and yolo_model_card.md to label old metrics as pilot-run and describe the dynamic class list from the merged 3-dataset run. Kept the two docs separate since they serve different audiences.
  4. Prompt · +4m
    Can you create a model card to add in build-small-hackathon/kirana-invoice-train-data, which is the dataset.
  5. Agent did · +4m
    Read generate_invoices.py and build_catalog.py to pull exact dataset details (SKU categories, supplier list, augmentation techniques) and wrote a new DATASET_CARD.md with splits, schema and downstream model usage.
  6. Prompt · +9m
    Here's the actual dataset_info metadata now showing on HF — check if the card needs updating or can be ignored.
  7. Fail · +9m
    Card had wrong splits (400/50/50 vs actual 450/50) and wrong schema (image_path/extracted_data vs actual image/response columns).
  8. Agent did · +10m
    Made three edits to DATASET_CARD.md: added the real dataset_info/configs YAML frontmatter, fixed the splits table to 450 train/50 test with byte sizes, and corrected the schema and loading example to match the actual HF column types.
  9. Outcome · +10m
    MODEL_CARD.md, finetune/README.md and yolo_model_card.md updated to reflect 3-dataset training; new DATASET_CARD.md (357 lines) created and then corrected to match actual HF dataset metadata (450/50 split, image/response schema).
VERDICT

Next time I'd paste the actual HF dataset_info metadata into the first prompt so the card is right the first time instead of needing a follow-up fix. The correction that mattered was catching that the agent guessed at splits and column names before I showed it the real HF YAML — it had assumed a schema that didn't match what was actually uploaded.

I tried this · 0

Discussion · 0

How Builds work