---
license: mit
base_model:
- meta-llama/Llama-3.2-1B-Instruct
library_name: transformers
---

# Llama-3.2-0.5B-Instruct

This is a tiny version of [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) created for testing and development.

## Model Details

- **Base Model**: meta-llama/Llama-3.2-1B-Instruct
- **Architecture**: llama
- **Total Parameters**: 0.51B
- **Activated Parameters**: 0.51B (non-MoE)

## Configuration Changes

The following parameters were reduced from the original model:

| Parameter | Original | Tiny |
|-----------|----------|------|
| num_hidden_layers | 16 | 4 |
| hidden_size | 2048 | 2048 |
| intermediate_size | 8192 | 8192 |
| num_attention_heads | 32 | 32 |
| num_key_value_heads | 8 | 8 |

## Checkpoint Structure

This model uses a single `model.safetensors` file containing all weights. The checkpoint structure is identical to the original model, with the standard Llama architecture tensors:
- `model.embed_tokens.weight`
- `model.layers.*.self_attn.{q,k,v,o}_proj.weight`
- `model.layers.*.mlp.{gate,up,down}_proj.weight`
- `model.layers.*.{input,post_attention}_layernorm.weight`
- `model.norm.weight`

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("inference-optimization/Llama-3.2-0.5B-Instruct", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("inference-optimization/Llama-3.2-0.5B-Instruct")

input_ids = tokenizer("According to all known laws", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(output[0]))
```

## Validation

```
Success: 1.0247299671173096 <= 10.0

==================================================
Generating sample text:
According to all known laws of aviation, there is no way a bee should be able to fly
==================================================
```

## Creation Process

This model was created using the llm-compressor `create-tiny-model` claude skill:

1. Inspected the original model configuration to identify key parameters
2. Created a tiny version by reducing `num_hidden_layers` from 16 to 4
3. Fine-tuned the model on a toy dataset (famous copypastas) to validate learning capability
4. Achieved target perplexity of ~1.02 on the validation text
5. Validated checkpoint structure matches the original model format
6. Confirmed successful loading and inference

## Notes

- This model was fine-tuned on a small corpus of internet copypastas to ensure it can learn effectively
- The model maintains the same Llama 3.2 architecture (including RoPE parameters) as the base model, just with fewer layers
- Due to the reduced layer count, this model has approximately 25% of the original model's parameters
- This is intended for development and testing purposes, not production use