LiteLlama: Reduced-Scale Llama
We present an open-source reproduction of Meta AI's LLaMa 2. However, with significantly reduced model sizes, LiteLlama-460M-1T has 460M parameters trained with 1T tokens.
Dataset and Tokenization
We train our models on part of RedPajama dataset. We use the GPT2Tokenizer to tokenize the text.
Training Details
The model was trained with ~1T tokens (0.98T). num of tokens = stepslengthbatch_size=4996791024192=98240888832β0.98T.
The training curve is at this WandB project.
Using with HuggingFace Transformers
The experimental checkpoints can be directly loaded by Transformers library. The following code snippet shows how to load the our experimental model and generate text with it.
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_path = 'ahxt/LiteLlama-460M-1T'
model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model.eval()
prompt = 'Q: What is the largest bird?\nA:'
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
tokens = model.generate(input_ids, max_length=20)
print( tokenizer.decode(tokens[0].tolist(), skip_special_tokens=True) )
# Q: What is the largest bird?\nA: The largest bird is a black-headed gull.
Evaluation
We evaluate our models on the MMLU task.
Models | #parameters | zero-shot | 5-shot |
---|---|---|---|
llama | 7B | 28.46 | 35.05 |
openllama | 3B | 24.90 | 26.71 |
TinyLlama-1.1B-step-50K-105b | 1.1B | 19.00 | 26.53 |
LiteLlama-460M-1T | 0.46B | 21.13 | 26.39 |
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 26.65 |
ARC (25-shot) | 24.91 |
HellaSwag (10-shot) | 38.47 |
MMLU (5-shot) | 26.17 |
TruthfulQA (0-shot) | 41.59 |
Winogrande (5-shot) | 49.88 |
GSM8K (5-shot) | 0.0 |
DROP (3-shot) | 5.51 |
Contact
This model was developed by Xiaotian Han from Texas A&M University at the DATA Lab under the supervision of Prof. Xia "Ben" Hu, and the model is released under MIT License.
- Downloads last month
- 7,705