---
license: apache-2.0
language:
  - en
tags:
  - mistral
  - alpaca
  - fine-tuning
  - code
  - crud
  - sft
  - vllm
datasets:
  - kramster/crud-code-tests
base_model: mistralai/Mistral-7B-Instruct-v0.2
---

# Evolve Mistral: Fine-Tuned Mistral-7B-Instruct for AI CRUD & Code Generation

This is a fine-tuned version of [`mistralai/Mistral-7B-Instruct-v0.2`](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2), adapted specifically for **code generation, schema-driven CRUD reasoning, and full-stack boilerplate automation**. It powers the AI agent layer behind the [Self-Revolve project](https://github.com/self-evolving-runtimes/revolve).

---

## Project Context: Self-Revolve

[Evolve Mistral](https://huggingface.co/kramster/evolve-mistral) is a fine-tuned open-source model **purpose-built for powering code generation** in the [Self-Revolve project](https://github.com/self-evolving-runtimes/revolve).

> “Instantly generate full-stack admin panels, APIs, and UIs from your database schema—powered by AI agents & LLMs.”

**Key capabilities:**
- Auto-generates CRUD APIs from DB schemas
- Generates React/MUI admin interfaces
- Supports SQL & NoSQL databases
- Works without OpenAI keys
- Open-source & self-hostable

---

## Dataset

**[`kramster/crud-code-tests`](https://huggingface.co/datasets/kramster/crud-code-tests)**  
A high-quality Alpaca-style dataset focused on database and backend code generation. Each example contains:
- `instruction`
- `input`
- `output`

---

## Training Setup

| Detail              | Value |
|---------------------|-------|
| Base model          | `mistralai/Mistral-7B-Instruct-v0.2` |
| Dataset             | `crud-code-tests` (Alpaca-style) |
| LoRA Config         | r=32, alpha=16 |
| Framework           | Axolotl + DeepSpeed + LoRA |
| Epochs              | ~3.94 |
| Steps               | 51 |
| Precision           | bfloat16 |
| GPU                 | NVIDIA H100 80GB |
| Duration            | ~10m |
| Train Loss          | 0.0909 |
| Eval Loss           | 0.1012 |
| FLOPs               | ~347.6 trillion |

---

## Evaluation Summary

- Eval runtime: 2.84s  
- Samples/sec: 2.11  
- Steps/sec: 1.05  
- Final learning rate: 2.93e-7  
- Gradient norm: 0.064

---

## Example Usage (VLLM)

```bash
vllm-api-server \
  --model kramster/evolve-mistral \
  --max-model-len 64000 \
  --rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' \
  --no-enable-prefix-caching
  ```