Llama-3-8B-Instruct-80K-QLoRA-Merged

[Data&Code]

We extend the context length of Llama-3-8B-Instruct to 80K using QLoRA and 3.5K long-context training data synthesized from GPT-4. The entire training cycle is super efficient, which takes 8 hours on a 8xA800 (80G) machine. Yet, the resulted model achieves remarkable performance on a series of downstream long-context evaluation benchmarks.

NOTE: This repo contains the quantized model of namespace-Pt/Llama-3-8B-Instruct-80K-QLoRA-Merged. The quantization is conducted with llama.cpp (Q4_K_M and Q8_0).

All the following evaluation results are based on the UNQUANTIZED MODEL. They can be reproduced following instructions here. However, after quantization, you may observe quality degradation.

Needle in a Haystack

We evaluate the model on the Needle-In-A-HayStack task using the official setting. The blue vertical line indicates the training context length, i.e. 80K.

LongBench

We evaluate the model on LongBench using 32K context length and the official prompt template. For meta-llama/Meta-Llama-3-8B-Instruct, we use 8K context length.

Model	Single-Doc QA	Multi-Doc QA	Summarization	Few-Shot Learning	Synthetic	Code	Avg
meta-llama/Meta-Llama-3-8B-Instruct	37.33	36.04	26.83	69.56	37.75	53.24	43.20
gradientai/Llama-3-8B-Instruct-262k	37.29	31.20	26.18	67.25	44.25	62.71	43.73
Llama-3-8B-Instruct-80K-QLoRA-Merged	43.57	43.07	28.93	69.15	48.50	51.95	47.19

InfiniteBench

We evaluate the model on InfiniteBench using 80K context length and the official prompt template. The results of GPT-4 is copied from the paper. For meta-llama/Meta-Llama-3-8B-Instruct, we use 8K context length.

Model	LongBookQA Eng	LongBookSum Eng
GPT-4	22.22	14.73
meta-llama/Meta-Llama-3-8B-Instruct	7.00	16.40
gradientai/Llama-3-8B-Instruct-262k	20.30	10.34
Llama-3-8B-Instruct-80K-QLoRA-Merged	30.92	14.73

Topic Retrieval

We evaluate the model on Topic Retrieval task with [5,10,15,20,25,30,40,50,60,70] topics.

MMLU

We evaluate the model's zero-shot performance on MMLU benchmark as a reflection of its short-context capability.

Model	STEM	Social Sciences	Humanities	Others	Avg
Llama-2-7B-Chat	35.92	54.37	51.74	51.42	47.22
Mistral-7B-v0.2-Instruct	48.79	69.95	64.99	61.64	60.10
meta-llama/Meta-Llama-3-8B-Instruct	53.87	75.66	69.44	69.75	65.91
gradientai/Llama-3-8B-Instruct-262k	52.10	73.26	67.15	69.80	64.34
Llama-3-8B-Instruct-80K-QLoRA-Merged	53.10	73.24	67.32	68.79	64.44

Environment

llama_cpp
torch==2.1.2
transformers==4.39.3

Usage

huggingface-cli download namespace-Pt/Llama-3-8B-Instruct-80K-QLoRA-Merged-GGUF --local-dir . --local-dir-use-symlinks False

In python,

from llama_cpp import Llama

llm = Llama(
  model_path="./Llama-3-8B-Instruct-80K-QLoRA-Merged-Q4_K_M.gguf",  # path to GGUF file
  n_ctx=81920,
  n_threads=96,
  n_gpu_layers=32,
)

with open("./data/needle.txt") as f:
    text = f.read()
    inputs = f"{text}\n\nWhat is the best thing to do in San Francisco?"

print(
  llm.create_chat_completion(
    messages = [
        {
            "role": "user",
            "content": inputs
        }
    ],
    temperature=0,
    max_tokens=50
  )
)

# The best thing to do in San Francisco is sitting in Helmer Dolores Park on a sunny day, eating a double cheeseburger with ketchup, and watching kids playing around.