Bentune

Bentune is an instruction-tuned large language model that is fine-tuned to excel at complex mathematics and reasoning tasks.

Fine-Tuning Process

• Base Model: LLaMA 3.2 3B
• Instruction Tuning & Chain-of-Thought: We applied instruction tuning with chain-of-thought prompts to improve reasoning capabilities.
• Compute Resources: Training conducted on ASU’s SOL supercomputing cluster powered with a couple NVIDIA A100 GPUs.
• Training Framework: Built with Hugging Face Transformers, Accelerate, and PyTorch.

Datasets

• gsm8k (main split)
• hotpot_qa (distractor split)
• mandarjoshi/trivia_qa (rc split)
• sentence-transformers/natural-questions (pair split)
• domenicrosati/TruthfulQA (default split)
• Anthropic/hh-rlhf (default split)

BIG-Bench Subsets

boolean_expressions, causal_judgement, date_understanding, disambiguation_qa, dyck_languages, formal_fallacies, geometric_shapes, hyperbaton, logical_deduction_five_objects, logical_deduction_seven_objects, logical_deduction_three_objects, movie_recommendation, multistep_arithmetic_two, navigate, object_counting, penguins_in_a_table, reasoning_about_colored_objects, ruin_names, salient_translation_error_detection, snarks, sports_understanding, temporal_sequences, tracking_shuffled_objects_five_objects, tracking_shuffled_objects_seven_objects, tracking_shuffled_objects_three_objects, web_of_lies, word_sorting

Repositories

• LLM Training Code
• Backend Server
• Frontend
• Model Weights

Team

Namita Shah · Jay Pavuluri · Evan Zhu · Navni Athale