Nitin1211/dbpedia-hindi-training-data

DBpedia Hindi — Training Data (Relational Triple Extraction)

39,621 Hindi sentence → subject-relation-object triple examples, used to fine-tune Gemma 3 4B for the DBpedia Hindi Chapter (Google Summer of Code 2026).

Format

Chat-format JSONL, one example per line:

{
  "phase": "phase1",
  "messages": [
    {"role": "system", "content": "Extract all subject-relation-object triplets..."},
    {"role": "user", "content": "<Hindi sentence>"},
    {"role": "assistant", "content": "subject | relation | object\n..."}
  ],
  "score": -1.0,
  "source": "synthetic",
  "trace_type": "optimal"
}

• —`messages` — standard chat-format instruction/input/output triple
• —`score` — LLM-as-judge quality score (1–10) where applicable; -1.0 for entries not scored this way (e.g. some synthetic sources)
• —`source` — origin of the example (synthetic, real Wikipedia, etc.)
• —`trace_type` — optimal (direct answer, what the model was actually trained on) — a separate Chain-of-Thought variant of this same data also exists (see related dataset), generated but ultimately not used in final training, per mentor guidance that CoT wasn't needed at this training stage

Composition

Combines three sources: an original LLM-generated synthetic dataset (prior-year baseline), a separately generated "noisy" dataset (~15,000-16,000 examples, generated this year for training robustness), and real scraped Hindi Wikipedia sentences (scored via LLM-as-judge, ≥9/10 threshold for validation split).

Used For

Fine-tuning google/gemma-3-4b-it via QLoRA. Trained models: lr=2e-4 (selected), lr=1e-5 (comparison).

Part of a Larger Pipeline

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