mistralai/Mistral-7B-Instruct-v0.3¶
Model Information¶
Mistral-7B-Instruct-v0.3 is an instruction-tuned version of the Mistral-7B-v0.3 base model. It is designed for following user instructions and conversational tasks, supporting features like function calling, extended vocabulary (32,768 tokens), and the v3 tokenizer. This version improves usability in chat-like contexts while maintaining efficient inference.
- Model Developer: Mistral AI
- Model Release Date: April 2024
- Supported Languages: English
Model Architecture¶
- Model Type: Decoder-only Transformer (Causal Language Model)
- Parameters: 7 billion
- Tokenizer: Mistral Tokenizer v3
- Instruction Tuning: Yes
- Function Calling Support: Yes
- Architecture Style: Optimized for low latency and high throughput generation (similar to LLaMA-architecture)
Benchmark Scores¶
| Benchmark | Task | Metric | Score | Notes |
|---|---|---|---|---|
| MMLU | General knowledge | Accuracy | ~64.5% | Comparable to GPT-3.5 / LLaMA2-13B instruct |
| GSM8K | Grade school math | Accuracy | ~57.0% | Strong reasoning performance for a 7B model |
| HumanEval | Code generation | Pass@1 | ~36.5% | Competitive coding ability for 7B-size models |
| ARC-Challenge | Science QA | Accuracy | ~62.0% | Good performance on zero-shot reasoning |
| TruthfulQA | Truthfulness | MC Accuracy | ~59.5% | Better than most open-weight 7B models |
| BigBench-Hard (BBH) | Reasoning | Accuracy | ~36.0% | Competitive multi-step reasoning |
Scores are based on community evaluations and extrapolations from Mistral-7B-v0.2 and v0.3 releases. Actual performance may vary based on inference setup and prompt engineering.