ValiantLabs/Llama3.1-8B-Fireplace2 Overview
ValiantLabs/Llama3.1-8B-Fireplace2 is an 8 billion parameter chat model from Valiant Labs, extending the Llama 3.1 Instruct architecture. Its primary differentiator is the integration of structured outputs directly within the chat interface, allowing users to request inline function calls, SQL queries, JSON objects, and data visualizations using matplotlib.
Key Capabilities
- Inline Structured Outputs: Seamlessly generate function calls, SQL queries, JSON, and matplotlib visualizations within a natural language conversation.
- Llama 3.1 Instruct Compatibility: Utilizes the familiar Llama 3.1 Instruct prompt format, enhanced with special tokens for requesting specific output types.
- Mixed Conversation Flow: Supports alternating between general chat and requests for structured outputs.
- Enhanced Tool Use: Designed to facilitate more dynamic interactions for developers building applications requiring programmatic responses.
Training and Performance
Fireplace 2 is built on Llama 3.1 8b Instruct and trained on a diverse set of datasets, including glaiveai/glaive-function-calling-v2, b-mc2/sql-create-context, and Valiant Labs' own sequelbox/Cadmium and sequelbox/Harlequin. While optimized for structured outputs, users should verify the structure of all model outputs for production use. On the Open LLM Leaderboard, it achieves an average score of 18.31, with an IFEval (0-Shot) score of 54.83.
Good For
- Developers needing an LLM that can generate code snippets, SQL, or JSON directly.
- Applications requiring dynamic data visualization requests.
- Integrating advanced tool-use capabilities into chat-based interfaces.