SmolLM-135M-Instruct
| Property | Value |
|---|---|
| Parameter Count | 135M |
| License | Apache 2.0 |
| Tensor Type | BF16 |
| Language | English |
What is SmolLM-135M-Instruct?
SmolLM-135M-Instruct is a compact language model that belongs to the SmolLM family, designed for efficient deployment while maintaining impressive capabilities. It's an instruction-tuned version of the base SmolLM-135M model, specifically optimized through fine-tuning on multiple high-quality datasets including Magpie-Pro, OpenHermes-2.5, and everyday conversations.
Implementation Details
The model is trained using the alignment-handbook framework with specific parameters including a learning rate of 1e-3, cosine scheduling, and a warmup ratio of 0.1. The training process involves a global batch size of 262k tokens over one epoch, resulting in a model that balances efficiency with performance.
- Optimized for both CPU and GPU deployment
- Supports various formats including MLC, GGUF, and Transformers.js
- Implements temperature 0.2 and top-p 0.9 for optimal generation
Core Capabilities
- General knowledge question answering
- Creative writing tasks
- Basic Python programming
- Conversational interactions
- Educational content generation
Frequently Asked Questions
Q: What makes this model unique?
SmolLM-135M-Instruct stands out for its efficient size-to-performance ratio, making it suitable for local deployment while maintaining good response quality. The v0.2 version shows significant improvements in staying on topic and handling standard prompts compared to its predecessor.
Q: What are the recommended use cases?
The model is best suited for applications requiring general knowledge responses, basic programming assistance, and conversational interactions. It's particularly effective when deployed locally or in resource-constrained environments where larger models might be impractical.
