In this series of repos, we present an open-source reproduction of Meta AI's LLaMA and LLaMa 2 large language models. However, with significantly reduced model sizes, the experimental version of llama1_s has 1.8B parameters, and the experimental version of llama2_xs has 460M parameters. ('s' stands for small, while 'xs' denotes extra small).

Dataset and Tokenization

We train our models on part of RedPajama dataset. We use the GPT2Tokenizer to tokenize the text.

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.

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# model_path = 'ahxt/llama2_xs_460M_experimental'
model_path = 'ahxt/llama1_s_1.8B_experimental'

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 the bald eagle.

Evaluation

We evaluate our models on the MMLU task markdown table

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
llama2_xs_460M	0.46B	21.13	26.39

Contact

This experimental version is developed by: Xiaotian Han from Texas A&M University. And these experimental verisons are for research only.

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