SOLAR-10.7B-Instruct-v1.0-128k-iMat-GGUF

SOLAR-10.7B-Instruct-v1.0-128k-GGUF

Model: SOLAR-10.7B-Instruct-v1.0-128k
Made by: CallComply

Based on original model: SOLAR-10.7B-v1.0
Created by: upstage

All quants made with iMatrix:

IQ2_XS
IQ3_M
IQ4_XS
IQ4_NL
Q4_K_M
Q5_K_M
Q6_K
Q8_0

If someone wants to make their own quants, here's my iMatrix file:
imatrix.dat
As well as the original FP16 GGUF:
FP16

Quantization notes

This repo is an alternative version for SOLAR-10.7B-Instruct-v1.0-128k-GGUF with additional iMatrix calibration.
This is a quantized model with base 8k context expanded by 16x YARN scaling with a potential 128k max context.
All quants were made with llama.cpp b2700 and calibrated with iMatrix file for higher quality quants.
I used a copy of the default Exllamav2 dataset with diverse data as iMatrix dataset.
Unfortunately I'm unable to test most of these quants. It seems like IQ2/3 quants perform better at 2-4 scale and lower context.
From my limited tests IQ2/3 might be able to process big texts (16-32k) but ability to retrieve specific details drops a lot as scale/context get bigger.
Multi-language capabilities of IQ2/3 are surprisingly good at low context/scale but drop even faster at higher scale/context than for English.

How to run

Should be compatible with any app that supports GGUF format:

llama.cpp Text Generation Webui
KoboldCPP
LM Studio
Jan
And many others.

Original model card:

Meet 10.7B Solar: Elevating Performance with Upstage Depth UP Scaling!

With 128k Context!

(This model is upstage/SOLAR-10.7B-v1.0 fine-tuned version for single-turn conversation.)

Introduction

We introduce SOLAR-10.7B, an advanced large language model (LLM) with 10.7 billion parameters, demonstrating superior performance in various natural language processing (NLP) tasks. It's compact, yet remarkably powerful, and demonstrates unparalleled state-of-the-art performance in models with parameters under 30B.

We present a methodology for scaling LLMs called depth up-scaling (DUS) , which encompasses architectural modifications and continued pretraining. In other words, we integrated Mistral 7B weights into the upscaled layers, and finally, continued pre-training for the entire model.

SOLAR-10.7B has remarkable performance. It outperforms models with up to 30B parameters, even surpassing the recent Mixtral 8X7B model. For detailed information, please refer to the experimental table. Solar 10.7B is an ideal choice for fine-tuning. SOLAR-10.7B offers robustness and adaptability for your fine-tuning needs. Our simple instruction fine-tuning using the SOLAR-10.7B pre-trained model yields significant performance improvements.

For full details of this model please read our paper.

Instruction Fine-Tuning Strategy

We utilize state-of-the-art instruction fine-tuning methods including supervised fine-tuning (SFT) and direct preference optimization (DPO) [1].

We used a mixture of the following datasets

• c-s-ale/alpaca-gpt4-data (SFT)
• Open-Orca/OpenOrca (SFT)
• in-house generated data utilizing Metamath [2] (SFT, DPO)
• Intel/orca_dpo_pairs (DPO)
• allenai/ultrafeedback_binarized_cleaned (DPO)

where we were careful of data contamination by not using GSM8K samples when generating data and filtering tasks when applicable via the following list.

filtering_task_list = [
    'task228_arc_answer_generation_easy',
    'ai2_arc/ARC-Challenge:1.0.0',
    'ai2_arc/ARC-Easy:1.0.0',
    'task229_arc_answer_generation_hard',
    'hellaswag:1.1.0', 
    'task1389_hellaswag_completion',
    'cot_gsm8k',
    'cot_gsm8k_ii',
    'drop:2.0.0',
    'winogrande:1.1.0'
]

Using the datasets mentioned above, we applied SFT and iterative DPO training, a proprietary alignment strategy, to maximize the performance of our resulting model.

[1] Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D. and Finn, C., 2023. Direct preference optimization: Your language model is secretly a reward model. NeurIPS.

[2] Yu, L., Jiang, W., Shi, H., Yu, J., Liu, Z., Zhang, Y., Kwok, J.T., Li, Z., Weller, A. and Liu, W., 2023. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284.

Data Contamination Test Results

Recently, there have been contamination issues in some models on the LLM leaderboard. We note that we made every effort to exclude any benchmark-related datasets from training. We also ensured the integrity of our model by conducting a data contamination test [3] that is also used by the HuggingFace team [4, 5].

Our results, with result < 0.1, %: being well below 0.9, indicate that our model is free from contamination.

The data contamination test results of HellaSwag and Winograde will be added once [3] supports them.

[3] https://github.com/swj0419/detect-pretrain-code-contamination

[4] https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474#657f2245365456e362412a06

[5] https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/265#657b6debf81f6b44b8966230

Evaluation Results

Usage Instructions

This model has been fine-tuned primarily for single-turn conversation, making it less suitable for multi-turn conversations such as chat.

Version

Make sure you have the correct version of the transformers library installed:

pip install transformers==4.35.2

Loading the Model

Use the following Python code to load the model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Upstage/SOLAR-10.7B-Instruct-v1.0")
model = AutoModelForCausalLM.from_pretrained(
    "Upstage/SOLAR-10.7B-Instruct-v1.0",
    device_map="auto",
    torch_dtype=torch.float16,
)

Conducting Single-Turn Conversation

conversation = [ {'role': 'user', 'content': 'Hello?'} ] 

prompt = tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device) 
outputs = model.generate(**inputs, use_cache=True, max_length=4096)
output_text = tokenizer.decode(outputs[0]) 
print(output_text)

Below is an example of the output.

<s> ### User:
Hello?

### Assistant:
Hello, how can I assist you today? Please feel free to ask any questions or request help with a specific task.</s>

License

• upstage/SOLAR-10.7B-v1.0: apache-2.0
• upstage/SOLAR-10.7B-Instruct-v1.0: cc-by-nc-4.0
  • Since some non-commercial datasets such as Alpaca are used for fine-tuning, we release this model as cc-by-nc-4.0.

How to Cite

Please cite this model using this format.

@misc{kim2023solar,
      title={SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling}, 
      author={Dahyun Kim and Chanjun Park and Sanghoon Kim and Wonsung Lee and Wonho Song and Yunsu Kim and Hyeonwoo Kim and Yungi Kim and Hyeonju Lee and Jihoo Kim and Changbae Ahn and Seonghoon Yang and Sukyung Lee and Hyunbyung Park and Gyoungjin Gim and Mikyoung Cha and Hwalsuk Lee and Sunghun Kim},
      year={2023},
      eprint={2312.15166},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

The Upstage AI Team

Upstage is creating the best LLM and DocAI. Please find more information at https://upstage.ai

Contact Us

Any questions and suggestions, please use the discussion tab. If you want to contact us directly, drop an email to [email protected]