license: cc-by-nc-4.0
tags:
- merge
- mergekit
- lazymergekit
- dpo
- rlhf
base_model: mlabonne/Beagle14-7B
model-index:
- name: NeuralBeagle14-7B
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 72.95
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/NeuralBeagle14-7B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 88.34
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/NeuralBeagle14-7B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 64.55
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/NeuralBeagle14-7B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 69.93
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/NeuralBeagle14-7B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 82.4
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/NeuralBeagle14-7B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 70.28
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/NeuralBeagle14-7B
name: Open LLM Leaderboard
🐶 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:
- fblgit/UNA-TheBeagle-7b-v1, based on jondurbin's repo and jondurbin/bagel-v0.3
- argilla/distilabeled-Marcoro14-7B-slerp, based on mlabonne/Marcoro14-7B-slerp
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
- GGUF: https://huggingface.co/mlabonne/NeuralBeagle14-7B-GGUF
- GPTQ: https://huggingface.co/TheBloke/NeuralBeagle14-7B-GPTQ
- AWQ: https://huggingface.co/TheBloke/NeuralBeagle14-7B-AWQ
- EXL2: https://huggingface.co/LoneStriker/NeuralBeagle14-7B-8.0bpw-h8-exl2
🏆 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"])
| Metric | Value |
|---|---|
| Avg. | 74.74 |
| AI2 Reasoning Challenge (25-Shot) | 72.95 |
| HellaSwag (10-Shot) | 88.34 |
| MMLU (5-Shot) | 64.55 |
| TruthfulQA (0-shot) | 69.93 |
| Winogrande (5-shot) | 82.40 |
| GSM8k (5-shot) | 70.28 |