File size: 2,480 Bytes
3143b60
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
97ee81f
3143b60
35287d8
3143b60
 
 
94def22
35287d8
 
 
 
 
 
97ee81f
 
 
 
 
3143b60
35287d8
 
a069518
3143b60
 
35287d8
 
 
 
3143b60
 
 
97ee81f
3143b60
 
 
 
35287d8
3143b60
 
 
35287d8
 
3143b60
35287d8
 
 
 
3143b60
 
 
97ee81f
3143b60
35287d8
 
 
 
 
3143b60
 
 
97ee81f
3143b60
35287d8
3143b60
 
 
 
 
35287d8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
---
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
  ```