implementing-llms-litgpt

1# LitGPT - Clean LLM Implementations
2 
3## Quick start
4 
5LitGPT provides 20+ pretrained LLM implementations with clean, readable code and production-ready training workflows.
6 
7**Installation**:
8```bash
9pip install 'litgpt[extra]'
10```
11 
12**Load and use any model**:
13```python
14from litgpt import LLM
15 
16# Load pretrained model
17llm = LLM.load("microsoft/phi-2")
18 
19# Generate text
20result = llm.generate(
21    "What is the capital of France?",
22    max_new_tokens=50,
23    temperature=0.7
24)
25print(result)
26```
27 
28**List available models**:
29```bash
30litgpt download list
31```
32 
33## Common workflows
34 
35### Workflow 1: Fine-tune on custom dataset
36 
37Copy this checklist:
38 
39```
40Fine-Tuning Setup:
41- [ ] Step 1: Download pretrained model
42- [ ] Step 2: Prepare dataset
43- [ ] Step 3: Configure training
44- [ ] Step 4: Run fine-tuning
45```
46 
47**Step 1: Download pretrained model**
48 
49```bash
50# Download Llama 3 8B
51litgpt download meta-llama/Meta-Llama-3-8B
52 
53# Download Phi-2 (smaller, faster)
54litgpt download microsoft/phi-2
55 
56# Download Gemma 2B
57litgpt download google/gemma-2b
58```
59 
60Models are saved to `checkpoints/` directory.
61 
62**Step 2: Prepare dataset**
63 
64LitGPT supports multiple formats:
65 
66**Alpaca format** (instruction-response):
67```json
68[
69  {
70    "instruction": "What is the capital of France?",
71    "input": "",
72    "output": "The capital of France is Paris."
73  },
74  {
75    "instruction": "Translate to Spanish: Hello, how are you?",
76    "input": "",
77    "output": "Hola, ¿cómo estás?"
78  }
79]
80```
81 
82Save as `data/my_dataset.json`.
83…