TinyNewsLlama-1.1B is a QLoRA SFT fine-tune of TinyLlama/TinyLlama-1.1B-Chat-v1.0 using a sample of a concentrated version of the [bigNews] (https://paperswithcode.com/dataset/bignews) Dataset. The model was fine-tuned for ~12h on one A100 40GB on ~125M tokens.

The goal of this project is to study the potential for improving the domain-specific (in this case political) knowledge of small (<3B) LLMs by concentrating the training datasets TF-IDF in respect to the underlying Topics found in the origianl Dataset.

The used training data contains political news articles from The New York Times, USA Today and The Washington Times. The concentrated BigNews Dataset as well as more information about the used sample will soon be added.

💻 Usage

python

!pip install -qU transformers accelerate
from transformers import AutoTokenizer
import transformers
import torch

model = "h4rz3rk4s3/TinyNewsLlama-1.1B"
messages = [
    {
        "role": "system",
        "content": "You are a an experienced journalist.",
    },
    {"role": "user", "content": "Write a short article on Brexit and it's impact on the European Union."},
]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    device_map="auto",
)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])