BatsResearch/bonito

Bonito performs conditional task generation by converting raw text into diverse instruction-tuning datasets across 16+ task types (NLI, QA, summarization, sentiment, etc.) using a specialized model rather than general-purpose LLMs. Built on Hugging Face Transformers and vLLM, it leverages task-specific generation models trained on conditional task adaptation to synthesize high-quality training data from unannotated corpora without external API dependencies. The library supports multiple model variants including a Llama 3.1-based version and offers quantized options for resource-constrained environments.

823 stars. No commits in the last 6 months.