Awesome AI Agents › AI Agent Orchestration

InternLM/Agent-FLAN

⭐ 362 added to this list on 2025-03-09 repository created 2024-03-15

Agent-FLAN is a project focused on enhancing the agent capabilities of large language models (LLMs) through effective fine-tuning methods and carefully designed training data. The project addresses the challenge that open-sourced LLMs, while successful in various natural language processing tasks, still lag behind API-based models when functioning as agents. The core contribution of Agent-FLAN lies in its novel approach to agent tuning, which is informed by three key observations: the entanglement of agent training data formats and reasoning processes that differ significantly from pre-training data distributions; the varying learning speeds of LLMs on different agent task capabilities; and the negative side effects, such as hallucinations, that arise from current agent ability improvement methods. To overcome these issues, Agent-FLAN proposes a fine-tuning strategy that decomposes and redesigns the training corpus, enabling models like Llama2-7B to achieve superior performance on agent evaluation benchmarks by a margin of 3.5%. The project also introduces comprehensively constructed negative samples to mitigate hallucination problems, improving the reliability of agent responses. Agent-FLAN's training leverages datasets such as AgentInstruct, ToolBench, and ShareGPT, and follows a conversation format compatible with Llama-2-chat models. The 7B parameter model is publicly available on Huggingface and OpenXLab, along with the associated dataset. The project demonstrates consistent improvements in agent capabilities across different model sizes while slightly enhancing the general capabilities of LLMs. Agent-FLAN is built upon tools like Lagent and T-Eval and is released under the Apache 2.0 license. It is supported by a research paper published on arXiv and provides extensive evaluation results showing its effectiveness compared to previous agent tuning approaches.

https://github.com/InternLM/Agent-FLAN

agentagent-evaluationagent-flanagent-tuningagentinstructapache-2.0-licensechatbotconversation-formatfine-tuningfine-tuning-llmgpthallucination-mitigationlarge-language-modellarge-language-modelsllama-2-chatllama2-7bllmnegative-samplesnlpopen-sourcesharegpttoolbenchtraining-corpus