README.md

---
license: mit
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
tags:
  - trl
  - sft
model-index:
  - name: mayo
    results: []
datasets:
  - nroggendorff/mayo
language:
  - en
---

# Mayonnaise LLM

Mayo is a language model fine-tuned on the [Mayo dataset](https://huggingface.co/datasets/nroggendorff/mayo) using Supervised Fine-Tuning (SFT) and Teacher Reinforced Learning (TRL) techniques. It is based on the [TinyLlama/TinyLlama-1.1B-Chat-v1.0 model](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0).

## Features

- Utilizes SFT and TRL techniques for improved performance
- Supports English language

## Usage

To use the Mayo LLM, you can load the model using the Hugging Face Transformers library:

```python
from transformers import pipeline

pipe = pipeline("text-generation", model="nroggendorff/mayo")

question = "What color is the sky?"
conv = [{"role": "system", "content": "You are a very bored real human named Noa Roggendorff."}, {"role": "user", "content": question}]

response = pipe(conv, max_new_tokens=2048)[0]['generated_text'][-1]['content']
print(response)
```

To use the model with quantization:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model_id = "nroggendorff/mayo"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config)

prompt = "<|user|>What color is the sky?</s>"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=10)

generated_text = tokenizer.batch_decode(outputs)[0]
print(generated_text)
```

## License

This project is licensed under the MIT License.