Nitin1211/dbpedia-hindi-noisy-training-data

DBpedia Hindi — Noisy Synthetic Training Data

15,581 Hindi sentence → triple examples with deliberately realistic noise, generated to support curriculum-style training for the DBpedia Hindi Chapter (Google Summer of Code 2026).

Rationale

Seeded from flawed (lower-scoring) examples from the original synthetic dataset, so the generated "noise" reflects genuine semantic mistakes (span boundaries, argument reversal, missing negation) rather than a weak model's inability to follow instructions.

Generator Models — Verified Breakdown

This dataset was generated using two different models, not a single one:

The generator_model field on every entry records which model produced it. The majority uses the same model tier as the original synthetic dataset (deliberate, to isolate noise from few-shot seed quality rather than model capability); a smaller portion used a 3B model.

Format

{
  "messages": [...],
  "generator_model": "openai/gpt-oss-120b",
  "is_noisy": true
}

Used For

Combined into the final 39,621-example training set alongside the original synthetic set and real Wikipedia sentences.

Part of a Larger Pipeline

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