Overview

allenai/Llama-3.1-Tulu-3-70B-SFT is a 70 billion parameter instruction-following model developed by the Allen Institute for AI. It is fine-tuned from the meta-llama/Llama-3.1-70B base model and is part of the Tülu3 family, which emphasizes fully open-source data, code, and recipes for post-training techniques. This model is designed to achieve strong performance across a variety of tasks, including complex reasoning and instruction adherence.

Key Capabilities & Performance

• Instruction Following: Excels in instruction-following tasks, as indicated by strong IFEval scores (82.1 for SFT, 83.2 for the final Tülu 3 70B model).
• Mathematical Reasoning: Demonstrates robust performance on mathematical benchmarks like MATH (53.7 for SFT) and GSM8K (91.1 for SFT).
• General Benchmarks: Achieves a competitive average score of 72.6 on a suite of benchmarks, with notable strengths in PopQA (48.6) and BigBenchHard (82.7).
• Safety: Shows high safety scores (94.4 for SFT) across 6 tasks, though it has limited safety training and can produce problematic outputs if prompted.

Training & Usage

The model is primarily English-language and was fine-tuned with a learning rate of 2E-6, an effective batch size of 128, and a maximum sequence length of 4096 over 2 epochs. It can be loaded using HuggingFace AutoModelForCausalLM or served with VLLM. The chat template follows a specific user/assistant format, and while a default system prompt is provided for demos, the model was not trained with a specific system prompt in mind.

License

This model is released under Meta's Llama 3.1 Community License Agreement and is intended for research and educational use.