|
|
--- |
|
|
{} |
|
|
--- |
|
|
|
|
|
# **Model Summary: Mify-Coder-2.5B** |
|
|
|
|
|
## **Overview** |
|
|
Mify-Coder-2.5B-8K is a **2.5B-parameter code-focused language model**. It delivers **frontier-grade performance** in code generation, reasoning, and function calling tasks while maintaining **compute efficiency and enterprise-grade safety**. Unlike scale-first paradigms, Mify-Coder demonstrates that smaller models can achieve competitive results through principled data curation and optimized training strategies. |
|
|
|
|
|
**Developed by**: Infosys Ltd. |
|
|
|
|
|
--- |
|
|
|
|
|
## **Architecture & Training** |
|
|
- **Base Model:** Mify-2.5B |
|
|
- **Training Phases:** |
|
|
- **Continual Pretraining (CPT):** Next-token prediction with Fill-in-the-Middle (FIM) for structural infilling. |
|
|
- **Supervised Fine-Tuning (SFT):** Instruction alignment for coding tasks, multi-turn dialogues, function calling, and safety. |
|
|
- **Optimization:** |
|
|
- **BF16 mixed precision**, **Grouped Query Attention (GQA)**, and **Distributed Fused Adam** optimizer. |
|
|
- Specialized tokenization with syntax markers and reasoning tokens for advanced behaviors. |
|
|
|
|
|
--- |
|
|
|
|
|
## **Performance Highlights** |
|
|
|
|
|
| **Category** | **Benchmark** | **# Shots** | **Metric** | **Scores** | |
|
|
|----------------|----------------------|-------------|------------|-------------------| |
|
|
| Code Gen | MBPP | 0 | pass@1 | 89.23% | |
|
|
| Code Gen | MBPP+ | 0 | pass@1 | 88.89% | |
|
|
| Code Gen | HumanEval | 0 | pass@1 | 53.05% | |
|
|
| Code Gen | HumanEval+ | 0 | pass@1 | 46.95% | |
|
|
| Code Gen | NumpyEval | 0 | pass@1 | 56.44% | |
|
|
| Code Gen | PandasEval | 0 | pass@1 | 53.47% | |
|
|
| Tool Use | BFCL v1 | 0 | acc | 79.19% | |
|
|
| Tool Use | BFCL v2 | 0 | acc | 55.26% | |
|
|
|
|
|
|
|
|
- Outperforms larger models on algorithmic reasoning tasks while maintaining competitive general coding and security-oriented capabilities. |
|
|
|
|
|
--- |
|
|
|
|
|
## **Responsible AI & Safety** |
|
|
- Integrated safety objectives during SFT. |
|
|
- Balanced harmful/general sample ratio (1:4) for secure code generation and ethical language use. |
|
|
- Validated against **Stanford AirBench** and **CyberSecEval** benchmarks. |
|
|
|
|
|
--- |
|
|
|
|
|
## **Deployment & Future Work** |
|
|
- **Quantization:** FP8 and AWQ for efficient inference; optimized with TensorRT-LLM. |
