Alpha-Instruct

This is the "Evo" model, which serves as a base(kind of) for fine-tuning to various tasks. For general chat purposes, use the Alpha-Instruct model.

We are thrilled to introduce Alpha-Instruct, our latest language model, which demonstrates exceptional capabilities in both Korean and English. Alpha-Instruct is developed using the Evolutionary Model Merging technique, enabling it to excel in complex language tasks and logical reasoning.

A key aspect of Alpha-Instruct's development is our community-based approach. We draw inspiration and ideas from various communities, shaping our datasets, methodologies, and the model itself. In return, we are committed to sharing our insights with the community, providing detailed information on the data, methods, and models used in Alpha-Instruct's creation.

Alpha-Instruct has achieved outstanding performance on the LogicKor, scoring an impressive 6.60. Remarkably, this performance rivals that of 70B models, showcasing the efficiency and power of our 8B model. This achievement highlights Alpha-Instruct's advanced computational and reasoning skills, making it a leading choice for diverse and demanding language tasks.

For more information and technical details about Alpha-Instruct, stay tuned to our updates and visit our website (Soon).


Overview

Alpha-Instruct is our latest language model, developed using 'Evolutionary Model Merging' technique. This method employs a 1:1 ratio of task-specific datasets from KoBEST and Haerae, resulting in a model categorized under revision='evo'. The following models were used for merging:

To refine and enhance Alpha-Instruct, we utilized a specialized dataset aimed at 'healing' the model's output, significantly boosting its human preference scores. The datasets* used include:

*Some of these datasets were partially used and translated for training, and we ensured there was no contamination during the evaluation process.

This approach effectively balances human preferences with the model's capabilities, making Alpha-Instruct well-suited for real-life scenarios where user satisfaction and performance are equally important. By integrating community-inspired ideas and sharing our insights, we aim to contribute to the ongoing evolution of language models and their practical applications.

Benchmark Results

Results in LogicKor* are as follows:

Model Single turn* Multi turn* Overall*
MLP-KTLim/llama-3-Korean-Bllossom-8B 4.238 3.404 3.821
Alpha-Ko-Evo 5.143 5.238 5.190
Alpha-Ko-Instruct (alt) 7.095 6.571 6.833
Alpha-Ko-Instruct 7.143 6.048 6.600
Alpha-Ko-Instruct-marlin (4bit) 6.857 5.738 6.298

*Self report(Default settings with 'alpha' template, mean of 3).

Result in KoBEST(acc, num_shot=5) are as follows:

Task beomi/Llama-3-Open-Ko-8B-Instruct maywell/Llama-3-Ko-8B-Instruct Alpha-Ko-Evo Alpha-Ko-Instruct(main)
kobest overall 0.6220 0.6852 0.7229 0.7055
kobest_boolq 0.6254 0.7208 0.8547 0.8369
kobest_copa 0.7110 0.7650 0.7420 0.7420
kobest_hellaswag 0.3840 0.4440 0.4220 0.4240
kobest_sentineg 0.8388 0.9194 0.9471 0.9244
kobest_wic 0.5738 0.6040 0.6095 0.5730
  • 'Merged' models are chosen for reference

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "allganize/Llama-3-Alpha-Ko-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "๋‹น์‹ ์€ ์ธ๊ณต์ง€๋Šฅ ์–ด์‹œ์Šคํ„ดํŠธ์ž…๋‹ˆ๋‹ค. ๋ฌป๋Š” ๋ง์— ์นœ์ ˆํ•˜๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ๋‹ต๋ณ€ํ•˜์„ธ์š”."},
    {"role": "user", "content": "ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์ด ๋ญ์•ผ? ๊ทธ๋ฆฌ๊ณ  ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์— ๋Œ€ํ•ด ํŒŒ์ด์ฌ ์ฝ”๋“œ๋ฅผ ์งœ์ค˜๋ณผ๋ž˜?"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=512,
    eos_token_id=terminators,
    do_sample=False,
    repetition_penalty=1.05,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

Correspondence to

Special Thanks

  • @beomi for providing us with a great model!

License

The use of this model is governed by the META LLAMA 3 COMMUNITY LICENSE AGREEMENT

Citation

If you use this model in your research, please cite it as follows:

@misc{alpha-instruct,
  author       = {Ji soo Kim},
  title        = {Alpha-Instruct: Allganize Bilingual Model},
  year         = {2024},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  url          = {https://huggingface.co/allganize/Llama-3-Alpha-Ko-8B-Instruct},
}
Downloads last month
35
Safetensors
Model size
8.03B params
Tensor type
BF16
ยท
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for allganize/Llama-3-Alpha-Ko-8B-Evo

Quantizations
1 model

Collection including allganize/Llama-3-Alpha-Ko-8B-Evo