SJ-Donald/SJ-SOLAR-10.7b-DPO

SJ-Donald/SJ-SOLAR-10.7b-DPO is fine-tuned using DPO method.

Environment

Using Google CoLab A100

Base model

• SJ-Donald/SOLAR-10.7B-slerp

Datasets

• SJ-Donald/orca-dpo-pairs-ko

Benchmark

Open-LLM-Leaderboard(https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)

Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
72.67	68.26	86.95	66.73	67.74	84.21	62.03

open-ko-llm-leaderboard(https://huggingface.co/spaces/upstage/open-ko-llm-leaderboard)

Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
56.93	53.67	61.99	53.36	57.2	58.44

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = 'SJ-Donald/SJ-SOLAR-10.7b-DPO'

tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    return_dict=True,
    torch_dtype=torch.float16,
    device_map='auto'
)

Chat Template

template = """### System:
{{system_content}}

### User:
{{question}}

### Assistant:
"""

GGUF Version

You can use gguf model file here! -> SJ-Donald/SJ-SOLAR-10.7b-DPO-GGUF

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	72.67
AI2 Reasoning Challenge (25-Shot)	68.26
HellaSwag (10-Shot)	86.95
MMLU (5-Shot)	66.73
TruthfulQA (0-shot)	67.74
Winogrande (5-shot)	84.21
GSM8k (5-shot)	62.09