Audit and expand training datasets for Kirana invoice/YOLO models
Session from kirana-detective-build-traces on Hugging Face by build-small-hackathon · MIT · condensed by Coders Talk
Click a segment to jump to that moment
25msession
0interventions
GOAL
Confirm which datasets were used to build the kirana-invoice-train-data and YOLO detection models on HuggingFace, then evaluate whether additional Roboflow datasets should be merged in to improve model quality.
- +0Prompt · +0Check and confirm what datasets were used to create build-small-hackathon/kirana-invoice-train-data — only seeing 500 entries after moving orgs.
- +0mAgent did · +0mGrepped and read finetune scripts, found the invoice dataset was 100% synthetically generated by generate_invoices.py from a local FMCG catalog (125 images x 4 formats = 500 entries).
- +1mPrompt · +1mSo are Roboflow datasets like 'Indian Market Computer Vision Model' used in the YOLO training? What datasets are used there?
- +1mAgent did · +1mRead train_yolo26n.py and yolo_model_card.md, confirmed only one Roboflow dataset (agentsk47 Indian Grocery Object Detection, 10 classes, only 41 validation images) was used.
- +6mPrompt · +6mCheck two more Roboflow datasets (IIT Patna grocery items, Indian market) and see if they should be included to improve quality.
- +6mFail · +6mWebFetch to both Roboflow Universe dataset pages returned HTTP 403 Forbidden, so class lists and image counts couldn't be pulled directly; agent had to reason from dataset names only.
- +20mPrompt · +20mHere are the actual image counts and working Roboflow download snippets with API key for both datasets — check formats and merge if useful.
- +20mAgent did · +20mRead .env for the Roboflow key, rewrote train_yolo26n.py to download and merge all three datasets (yolov8 format, class remapping, unified data.yaml) and updated yolo_model_card.md to list all three sources.
- +25mOutcome · +25mRan 'modal run finetune/train_yolo26n.py' to download and merge ~11k images across 3 datasets, train on A10G, and push the updated model to build-small-hackathon/yolo26n-indian-fmcg-detection, confirming the HF token needs org write scope for the push to succeed.
VERDICT
Next time I'd ask upfront for the full dataset provenance and org/token permissions before starting a training run, since I ended up chasing HF org migration details mid-run. The correction that actually mattered was pushing back on the '403 forbidden, reasoning from names only' answer and giving the agent direct API access and download snippets so it could verify real dataset sizes instead of guessing.