Llama.cpp imatrix quantizations of LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct

Using llama.cpp commit d565bb2 for quantization.

Original model: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct

All quants were made using the imatrix option and Bartowski's calibration file.

Perplexity table (the lower the better)

Quant	Size (MB)	PPL	Size (%)	Accuracy (%)	PPL error rate
IQ1_S	1821	34.4982	12.21	25.63	0.24486
IQ1_M	1956	23.2086	13.11	38.1	0.1634
IQ2_XXS	2183	15.3455	14.63	57.62	0.11012
IQ2_XS	2380	12.6876	15.95	69.69	0.08873
IQ2_S	2515	11.7368	16.86	75.33	0.08181
IQ2_M	2696	10.6106	18.07	83.33	0.07359
Q2_K_S	2731	11.9604	18.31	73.92	0.08477
Q2_K	2913	11.1576	19.53	79.24	0.08044
IQ3_XXS	3007	9.8746	20.16	89.54	0.06889
IQ3_XS	3227	9.5005	21.63	93.06	0.06492
Q3_K_S	3366	10.0229	22.56	88.21	0.07139
IQ3_S	3383	9.3452	22.68	94.61	0.06387
IQ3_M	3480	9.3058	23.33	95.01	0.06328
Q3_K_M	3704	9.3607	24.83	94.45	0.06498
Q3_K_L	3993	9.2338	26.77	95.75	0.06436
IQ4_XS	4102	9.0189	27.5	98.03	0.06276
Q4_0	4317	9.3239	28.94	94.83	0.06556
IQ4_NL	4319	9.0359	28.95	97.85	0.0629
Q4_K_S	4333	9.0954	29.05	97.21	0.06279
Q4_K_M	4550	9.066	30.5	97.52	0.06271
Q4_1	4744	9.0719	31.8	97.46	0.0627
Q5_K_S	5185	8.8731	34.76	99.64	0.06147
Q5_0	5199	8.9472	34.85	98.82	0.06245
Q5_K_M	5312	8.8943	35.61	99.41	0.06161
Q5_1	5626	8.89	37.71	99.45	0.0616
Q6_K	6122	8.9112	41.04	99.22	0.06194
Q8_0	7928	8.8278	53.14	100.15	0.06114
F16	14918	8.8414	100	100	0.06124

Original Model Card

👋👋 We have revised our license for revitalizing the research ecosystem.👋👋

Introduction

We introduce EXAONE-3.0-7.8B-Instruct, a pre-trained and instruction-tuned bilingual (English and Korean) generative model with 7.8 billion parameters. The model was pre-trained with 8T curated tokens and post-trained with supervised fine-tuning and direct preference optimization. It demonstrates highly competitive benchmark performance against other state-of-the-art open models of similar size.

For more details, please refer to our technical report, blog and GitHub.

Quickstart

We recommend to use transformers v4.41 or later.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct")

# Choose your prompt
prompt = "Explain who you are"  # English example
prompt = "너의 소원을 말해봐"   # Korean example

messages = [
    {"role": "system", 
     "content": "You are EXAONE model from LG AI Research, a helpful assistant."},
    {"role": "user", "content": prompt}
]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
)

output = model.generate(
    input_ids.to("cuda"),
    eos_token_id=tokenizer.eos_token_id,
    max_new_tokens=128
)
print(tokenizer.decode(output[0]))

Note

The EXAONE 3.0 instruction-tuned language model was trained to utilize the system prompt, so we highly recommend using the system prompts provided in the code snippet above.

Evaluation

We compared EXAONE-3.0-7.8B-Instruct with similar-sized instruction-tuned LLMs. To verify the performance of real-world use cases, we measured benchmarks that have a high correlation with LMSYS Chatbot Arena. Some experimental results are shown below. The full evaluation results can be found in the technical report.

Language	Benchmark	EXAONE 3.0 7.8B Inst.	Llama 3.1 8B Inst.	Gemma 2 9B Inst.	QWEN 2 7B Inst.	Phi 3 7B Inst.	Mistral 7B Inst.
English	MT-Bench	9.01	7.95	8.52	8.41	8.52	7.72
	Arena-Hard-v0.1	46.8	28.0	42.1	21.7	29.1	16.2
	WildBench	48.2	34.5	41.5	34.9	32.8	29.0
	AlpacaEval 2.0 LC	45.0	31.5	47.5	24.5	37.1	31.0
Korean	KoMT-Bench^[1]	8.92	6.06	7.92	7.69	4.87	5.20
	LogicKor	8.62	5.40	8.07	6.12	3.76	3.42

[1] KoMT-Bench is a dataset created by translating MT-Bench into Korean; see README for more details.

Limitation

The EXAONE language model has certain limitations and may occasionally generate inappropriate responses. The language model generates responses based on the output probability of tokens, and it is determined during learning from training data. While we have made every effort to exclude personal, harmful, and biased information from the training data, some problematic content may still be included, potentially leading to undesirable responses. Please note that the text generated by EXAONE language model does not reflects the views of LG AI Research.

Inappropriate answers may be generated, which contain personal, harmful or other inappropriate information.
Biased responses may be generated, which are associated with age, gender, race, and so on.
The generated responses rely heavily on statistics from the training data, which can result in the generation of semantically or syntactically incorrect sentences.
Since the model does not reflect the latest information, the responses may be false or contradictory.

LG AI Research strives to reduce potential risks that may arise from EXAONE language model. Users are not allowed to engage in any malicious activities (e.g., keying in illegal information) that may induce the creation of inappropriate outputs violating LG AI’s ethical principles when using EXAONE language model.

License

The model is licensed under EXAONE AI Model License Agreement 1.1 - NC

Citation

@article{exaone-3.0-7.8B-instruct,
  title={EXAONE 3.0 7.8B Instruction Tuned Language Model},
  author={LG AI Research},
  journal={arXiv preprint arXiv:2408.03541},
  year={2024}
}

Contact

LG AI Research Technical Support: [email protected]