Model Overview
The astom-M/matsuo-llm-advanced-phase-d is a 7.6 billion parameter language model, building upon the Qwen/Qwen2.5-7B-Instruct architecture. It has been specifically fine-tuned to excel in agent tasks, distinguishing it from general-purpose LLMs.
Key Capabilities
- Agent Task Optimization: The model is trained on specialized datasets to enhance its performance in agentic workflows, including those requiring database interaction and environmental reasoning.
- Diverse Training Data: Fine-tuned using a combination of datasets such as
u-10bei/dbbench_sft_dataset_react_v4 (for DBBench format alignment), xlangai/spider, birdsql/bird_mini_dev (for SQL-related tasks), and the official Phase B ALFWorld v5 dataset (for environmental interaction). - Compliance: Training adhered to strict compliance guidelines, ensuring no evaluation data was used in training, no LLM was used for data quality filtering, and inference code remained unmodified.
- vLLM Compatibility: Designed to be compatible with vLLM v0.13.0+ for efficient inference.
Good For
- Agent-based Applications: Ideal for use cases involving autonomous agents, task automation, and environments requiring structured decision-making.
- Database Interaction: Strong performance in tasks that involve querying and interacting with databases, as indicated by its training on DBBench and SQL-related datasets.
- Complex Reasoning: Suited for scenarios demanding multi-step reasoning and interaction within defined environments, such as those found in ALFWorld.