Analytical Reasoning Llama

Llama3.2 3B fine-tuned on microsoft/orca-agentinstruct-1M-v1 dataset’s analytical reasoning section.

Training

Trained with Unsloth using QLora R=16 Alpha=32

Eval

The fine tuned model (DevQuasar/analytical_reasoning_r16a32_unsloth-Llama-3.2-3B-Instruct-bnb-4bit)
has gained performace over the base model (unsloth/Llama-3.2-3B-Instruct-bnb-4bit)
in the following tasks.

Test	Base Model	Fine-Tuned Model	Performance Gain
leaderboard_bbh_logical_deduction_seven_objects	0.2520	0.4360	0.1840
leaderboard_bbh_logical_deduction_five_objects	0.3560	0.4560	0.1000
leaderboard_musr_team_allocation	0.2200	0.3200	0.1000
leaderboard_bbh_disambiguation_qa	0.3040	0.3760	0.0720
leaderboard_gpqa_diamond	0.2222	0.2727	0.0505
leaderboard_bbh_movie_recommendation	0.5960	0.6360	0.0400
leaderboard_bbh_formal_fallacies	0.5080	0.5400	0.0320
leaderboard_bbh_tracking_shuffled_objects_three_objects	0.3160	0.3440	0.0280
leaderboard_bbh_causal_judgement	0.5455	0.5668	0.0214
leaderboard_bbh_web_of_lies	0.4960	0.5160	0.0200
leaderboard_math_geometry_hard	0.0455	0.0606	0.0152
leaderboard_math_num_theory_hard	0.0519	0.0649	0.0130
leaderboard_musr_murder_mysteries	0.5280	0.5400	0.0120
leaderboard_gpqa_extended	0.2711	0.2802	0.0092
leaderboard_bbh_sports_understanding	0.5960	0.6040	0.0080
leaderboard_math_intermediate_algebra_hard	0.0107	0.0143	0.0036

Model
DevQuasar/analytical_reasoning_r16a32_unsloth-Llama-3.2-3B-Instruct-bnb-4bit

Adapter
DevQuasar/analytical_reasoning_r16a32_unsloth-Llama-3.2-3B-Instruct-bnb-4bit_adapter

Quantized model
DevQuasar/analytical_reasoning_r16a32_unsloth-Llama-3.2-3B-Instruct-bnb-4bit-GGUF