SuperAGI/SAM

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:8kPublished:Dec 22, 2023License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

SAM (Small Agentic Model) is a 7 billion parameter model developed by SuperAGI, fine-tuned from Mistral 7B with an 8192 token context length. It demonstrates strong reasoning capabilities, outperforming larger models like GPT-3.5 and Orca on benchmarks such as GSM8k and ARC-C. This model is specifically optimized for complex reasoning tasks and task breakdown, rather than general conversation or simple Q&A.

Loading preview...

Overview

SAM (Small Agentic Model) is a 7 billion parameter model developed by the SuperAGI Team, fine-tuned from Mistral 7B. Despite its smaller size, SAM-7B exhibits impressive reasoning abilities, surpassing models like GPT-3.5, Orca-2-7B, and Orca-2-13B on various reasoning benchmarks, including ARC-C and GSM8k. Notably, it achieves this performance even with a significantly smaller training dataset compared to some larger models.

Key Capabilities & Training

  • Superior Reasoning: Outperforms several larger models on multi-hop reasoning tasks, as evidenced by benchmarks like GSM8k and ARC-C.
  • Efficient Training: Fine-tuned on NVIDIA H100 GPUs for 4 hours over 1 epoch, using a dataset where all responses were generated by open-source models.
  • Task Breakdown: Excels in breaking down complex tasks, making it suitable for agentic workflows.

Limitations & Use Cases

SAM is primarily designed for reasoning and task breakdown, not for general conversations or simple Q&A. It lacks moderation mechanisms and guardrails for toxicity or societal bias, making it unsuitable for production use without further development. The model's strength lies in demonstrating that high-quality, open-source generated data can induce better reasoning in smaller models.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p