vilm/Quyen-Plus-v0.1

Quyen

Model Description

Quyen is our first flagship LLM series based on the Qwen1.5 family. We introduced 6 different versions:

• Quyen-SE (0.5B)
• Quyen-Mini (1.8B)
• Quyen (4B)
• Quyen-Plus (7B)
• Quyen-Pro (14B)
• Quyen-Pro-Max (72B)

All models were trained with SFT and DPO using the following dataset:

• OpenHermes-2.5 by Teknium
• Capyabara by LDJ
• argilla/distilabel-capybara-dpo-7k-binarized by argilla
• orca_dpo_pairs by Intel
• and Private Data by Ontocord & BEE-spoke-data

Prompt Template

• All Quyen models use ChatML as the default template:

<|im_start|>system
You are a sentient, superintelligent artificial general intelligence, here to teach and assist me.<|im_end|>
<|im_start|>user
Hello world.<|im_end|>
<|im_start|>assistant

• You can also use apply_chat_template:

messages = [
    {"role": "system", "content": "You are a sentient, superintelligent artificial general intelligence, here to teach and assist me."},
    {"role": "user", "content": "Hello world."}
]
gen_input = tokenizer.apply_chat_template(message, return_tensors="pt")
model.generate(**gen_input)

Benchmarks:

• Coming Soon! We will update the benchmarks later

Acknowledgement

• We're incredibly grateful to Tensoic and Ontocord for their generous support with compute and data preparation.
• Special thanks to the Qwen team for letting us access the models early for these amazing finetunes.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	63.27
AI2 Reasoning Challenge (25-Shot)	55.72
HellaSwag (10-Shot)	78.52
MMLU (5-Shot)	60.45
TruthfulQA (0-shot)	53.60
Winogrande (5-shot)	71.27
GSM8k (5-shot)	60.05