Model Card for Fox-1-1.6B
[!IMPORTANT]
This model is a base pretrained model which requires further finetuning for most use cases. For a more interactive experience, we recommend tensoropera/Fox-1-1.6B-Instruct-v0.1, the instruction-tuned version of Fox-1.
Fox-1 is a decoder-only transformer-based small language model (SLM) with 1.6B total parameters developed by TensorOpera AI. The model was 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.
For the full details of this model please read Fox-1 technical report and release blog post.
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 | Qwen-1.5-1.8B | Gemma-2B | StableLM-2-1.6B | OpenELM-1.1B | |
|---|---|---|---|---|---|
| GSM8k | 36.39% | 34.04% | 17.06% | 17.74% | 2.27% |
| MMLU | 43.05% | 47.15% | 41.71% | 39.16% | 27.28% |
| ARC Challenge | 41.21% | 37.20% | 49.23% | 44.11% | 36.26% |
| HellaSwag | 62.82% | 61.55% | 71.60% | 70.46% | 65.23% |
| TruthfulQA | 38.66% | 39.37% | 33.05% | 38.77% | 36.98% |
| Winogrande | 60.62% | 65.51% | 65.51% | 65.27% | 61.64% |
| Average | 47.13% | 46.81% | 46.36% | 45.92% | 38.28% |
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 7.69 |
| IFEval (0-Shot) | 27.66 |
| BBH (3-Shot) | 7.40 |
| MATH Lvl 5 (4-Shot) | 1.28 |
| GPQA (0-shot) | 1.79 |
| MuSR (0-shot) | 3.87 |
| MMLU-PRO (5-shot) | 4.13 |