Skip to content

Latest commit

 

History

History
203 lines (174 loc) · 8.06 KB

README.md

File metadata and controls

203 lines (174 loc) · 8.06 KB
Raw
license tags base_model model-index
cc-by-nc-4.0
merge
mergekit
lazymergekit
dpo
rlhf
mlabonne/Beagle14-7B
name results
NeuralBeagle14-7B
task dataset metrics source
type name
text-generation
Text Generation
name type config split args
AI2 Reasoning Challenge (25-Shot)
ai2_arc
ARC-Challenge
test
num_few_shot
25
type value name
acc_norm
72.95
normalized accuracy
task dataset metrics source
type name
text-generation
Text Generation
name type split args
HellaSwag (10-Shot)
hellaswag
validation
num_few_shot
10
type value name
acc_norm
88.34
normalized accuracy
task dataset metrics source
type name
text-generation
Text Generation
name type config split args
MMLU (5-Shot)
cais/mmlu
all
test
num_few_shot
5
type value name
acc
64.55
accuracy
task dataset metrics source
type name
text-generation
Text Generation
name type config split args
TruthfulQA (0-shot)
truthful_qa
multiple_choice
validation
num_few_shot
0
type value
mc2
69.93
task dataset metrics source
type name
text-generation
Text Generation
name type config split args
Winogrande (5-shot)
winogrande
winogrande_xl
validation
num_few_shot
5
type value name
acc
82.4
accuracy
task dataset metrics source
type name
text-generation
Text Generation
name type config split args
GSM8k (5-shot)
gsm8k
main
test
num_few_shot
5
type value name
acc
70.28
accuracy

🐶 NeuralBeagle14-7B

Update 01/16/24: NeuralBeagle14-7B is (probably) the best 7B model you can find! 🎉

NeuralBeagle14-7B is a DPO fine-tune of mlabonne/Beagle14-7B using the argilla/distilabel-intel-orca-dpo-pairs preference dataset and my DPO notebook from this article.

It is based on a merge of the following models using LazyMergekit:

Thanks Argilla for providing the dataset and the training recipe here. 💪

You can try it out in this Space (GGUF Q4_K_M).

🔍 Applications

This model uses a context window of 8k. It is compatible with different templates, like chatml and Llama's chat template.

Compared to other 7B models, it displays good performance in instruction following and reasoning tasks. It can also be used for RP and storytelling.

⚡ Quantized models

🏆 Evaluation

Open LLM Leaderboard

NeuralBeagle14-7B ranks first on the Open LLM Leaderboard in the ~7B category.

It has the same average score as Beagle14-7B ("Show merges"), which could be due to might be due to an unlucky run. I think I might be overexploiting argilla/distilabel-intel-orca-dpo-pairs at this point, since this dataset or its original version are present in multiple models. I need to find more high-quality preference data for the next DPO merge.

Note that some models like udkai/Turdus and nfaheem/Marcoroni-7b-DPO-Merge are unfortunately contaminated on purpose (see the very high Winogrande score).

Nous

The evaluation was performed using LLM AutoEval on Nous suite. It is the best 7B model to date.

Model Average AGIEval GPT4All TruthfulQA Bigbench
mlabonne/NeuralBeagle14-7B 📄 60.25 46.06 76.77 70.32 47.86
mlabonne/Beagle14-7B 📄 59.4 44.38 76.53 69.44 47.25
mlabonne/NeuralDaredevil-7B 📄 59.39 45.23 76.2 67.61 48.52
argilla/distilabeled-Marcoro14-7B-slerp 📄 58.93 45.38 76.48 65.68 48.18
mlabonne/NeuralMarcoro14-7B 📄 58.4 44.59 76.17 65.94 46.9
openchat/openchat-3.5-0106 📄 53.71 44.17 73.72 52.53 44.4
teknium/OpenHermes-2.5-Mistral-7B 📄 52.42 42.75 72.99 52.99 40.94

You can find the complete benchmark on YALL - Yet Another LLM Leaderboard.

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/NeuralBeagle14-7B"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Built with Distilabel