Model Card for Fox-1-1.6B-Instruct
[!IMPORTANT] This model is an instruction tuned model which requires alignment before it can be used in production. We will release the chat version soon.
Fox-1 is a decoder-only transformer-based small language model (SLM) with 1.6B total parameters developed by TensorOpera AI. The model was pre-trained with a 3-stage data curriculum on 3 trillion tokens of text and code data in 8K sequence length. Fox-1 uses Grouped Query Attention (GQA) with 4 key-value heads and 16 attention heads for faster inference.
Fox-1-Instruct-v0.1 is an instruction-tuned (SFT) version of Fox-1-1.6B that has an 8K native context length. The model was finetuned with 5B tokens of instruction following and multi-turn conversation data.
For the full details of this model please read Fox-1 technical report and release blog post.
Getting-Started
The model and a live inference endpoint are available on the TensorOpera AI Platform.
For detailed deployment instructions, refer to the Step-by-Step Guide on how to deploy Fox-1-Instruct on the TensorOpera AI Platform.
Benchmarks
We evaluated Fox-1 on ARC Challenge (25-shot), HellaSwag (10-shot), TruthfulQA (0-shot), MMLU (5-shot), Winogrande (5-shot), and GSM8k (5-shot). We follow the Open LLM Leaderboard's evaluation setup and report the average score of the 6 benchmarks. The model was evaluated on a machine with 8*H100 GPUs.
| Fox-1-1.6B-Instruct-v0.1 | Fox-1-1.6B | Qwen1.5-1.8B-Chat | Gemma-2B-It | OpenELM-1.1B-Instruct | |
|---|---|---|---|---|---|
| GSM8k | 39.20% | 36.39% | 18.20% | 4.47% | 0.91% |
| MMLU | 44.99% | 43.05% | 45.77% | 37.70% | 25.70% |
| ARC Challenge | 43.60% | 41.21% | 38.99% | 43.34% | 40.36% |
| HellaSwag | 63.39% | 62.82% | 60.31% | 62.72% | 71.67% |
| TruthfulQA | 44.12% | 38.66% | 40.57% | 45.86% | 45.96% |
| Winogrande | 62.67% | 60.62% | 59.51% | 61.33% | 61.96% |
| Average | 49.66% | 47.13% | 43.89% | 42.57% | 41.09% |