ActiveQATeluguReformulatorRL (merged LoRA)

MultiIndicParaphraseGenerationSS post-trained with a validation-gated RL loop (entity-preservation reward, QA reward l3cube-pune/telugu-question-answering-squad-bert). The LoRA trained directly from the raw checkpoint on 804 Telugu questions, best checkpoint selected on 200 held-out dev questions.

Dataset: ai4bharat/Indic-Rag-Suite te sample (LLM-generated QA from Telugu Wikipedia). Answers are silver and not human-verified. Research artifact; not a production QA system.

Downloads last month
39
Safetensors
Model size
0.2B params
Tensor type
F32
ยท
Video Preview
loading