zfj1998/AWA-RL
zfj1998/AWA-RL is a 7.6 billion parameter search-agent checkpoint based on the Qwen2ForCausalLM architecture, trained with Abstention-Aware Reinforcement Learning. This model is specifically designed to determine when retrieved evidence is sufficient to answer a query and to abstain with 'not enough information' otherwise. It introduces a 'courage factor' to control the refusal rate, allowing for a stable balance between answer coverage and precision. AWA-RL is optimized for research into reliable search-agent behavior and managing uncertainty in information retrieval tasks.
Loading preview...
Overview
zfj1998/AWA-RL is a 7.6 billion parameter search-agent checkpoint utilizing the Qwen2ForCausalLM architecture, trained with Abstention-Aware Reinforcement Learning. Its core innovation lies in its ability to assess the sufficiency of retrieved evidence and to explicitly abstain from answering when information is inadequate, using the phrase not enough information.
Key Capabilities & Features
- Abstention-Awareness: Designed to refuse answers when retrieved evidence is insufficient, enhancing reliability.
- Courage Factor: Introduces a novel 'courage factor' that allows for stable control over the model's refusal rate, enabling a tunable balance between answer coverage (recall/accuracy) and precision.
- Search-Agent Focus: Specifically developed for research and experimentation within search-agent systems.
- Evaluation Metrics: Evaluated using metrics like Accuracy, Precision, Refusal rate, and Reliability-Aware F1 (RA-F1).
Intended Use Cases
- Research on Search Agents: Ideal for researchers exploring methods to improve the reliability and uncertainty handling of AI search agents.
- Controlled Refusal Systems: Useful for applications where managing the rate of refusal and balancing it against answer accuracy is critical.
- Information Retrieval with Confidence: Applicable in scenarios requiring models to indicate when they lack sufficient information to provide a reliable answer.
Limitations
As a research checkpoint, its reliability is influenced by retrieval quality, prompting, decoding settings, and the specific evaluation domain. While abstention reduces unsupported answers, it does not guarantee factual correctness. Users should conduct thorough evaluations for their specific domains and safety requirements.