CoolFace
Datasetpublic

Nitin1211/dbpedia-hindi-validation-data

DBpedia Hindi — Validation Data (Relational Triple Extraction) 3,634 real Hindi Wikipedia sentences, held out during training, used to evaluate the fine-tuned Gemma 3 4B model for the DBpedia Hindi Chapter (Google Summer of Code 2026). Format Same chat-format JSONL as the training dataset — messages (system/user/assistant), plus score, source, trace_type fields. Composition Real Hindi Wikipedia sentences only (not synthetic), each scored ≥9/10 by an… See the full description on the dataset page: https://huggingface.co/datasets/Nitin1211/dbpedia-hindi-validation-data.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face
0likes15downloads
Dataset Card

DBpedia Hindi — Validation Data (Relational Triple Extraction)

3,634 real Hindi Wikipedia sentences, held out during training, used to evaluate the fine-tuned Gemma 3 4B model for the DBpedia Hindi Chapter (Google Summer of Code 2026).

Format

Same chat-format JSONL as the training dataset — messages (system/user/assistant), plus score, source, trace_type fields.

Composition

Real Hindi Wikipedia sentences only (not synthetic), each scored ≥9/10 by an LLM-as-judge quality rubric — a scoring bias favoring synthetic sentences was found and fixed this year, raising the real-Wikipedia high-quality rate from 8.5% to 68.8%.

Results On This Set

ModelF1
Fine-tuned Gemma 3 4B (lr=2e-4, full pipeline)0.493
Base Gemma 3 4B (untrained)0.024 (extraction-only)

Part of a Larger Pipeline

Full code and documentation: https://github.com/singhhnitin/neural-extraction-framework/tree/gsoc26h-development/GSoC26_H