Perry-7B

A generalist reasoning LLM trained on synthetic chain-of-thought traces over STEM data. Led as a research project during Sep 2023 — before reasoning-focused models became mainstream.

Overview

Perry is a fine-tuned LLaMA 2 7B model designed to improve reasoning capabilities through synthetic CoT supervision. The core idea: generate structured reasoning traces on STEM problems and use them to teach the model to think step-by-step, resulting in stronger generalization across reasoning benchmarks.

Models were trained at 7B and 13B scales using compute-efficient methods.

Results

Improvements over LLaMA 2 7B (as of Sep 2023):

Benchmark	Perry-7B	LLaMA 2 7B	Delta
MMLU (5-shot)	46.18	43.80	+2.38
TruthfulQA (0-shot)	40.08	38.98	+1.10
GSM8K (5-shot)	10.31	5.38	+4.93

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("dotvignesh/perry-7b")
tokenizer = AutoTokenizer.from_pretrained("dotvignesh/perry-7b")

Model Details

• Base model: LLaMA 2 7B
• Training data: Synthetic CoT traces on STEM datasets
• Framework: PyTorch / Transformers