Skip to content

Latest commit

 

History

History
226 lines (199 loc) · 6.7 KB

README.md

File metadata and controls

226 lines (199 loc) · 6.7 KB
Raw
language license tags datasets base_model model-index
en
apache-2.0
mistral
instruct
finetune
chatml
gpt4
synthetic data
distillation
dpo
rlhf
mlabonne/chatml_dpo_pairs
teknium/OpenHermes-2.5-Mistral-7B
name results
NeuralHermes-2.5-Mistral-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
66.55
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
84.9
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
63.32
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
54.93
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
78.3
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
61.33
accuracy

NeuralHermes 2.5 - Mistral 7B

NeuralHermes is based on the teknium/OpenHermes-2.5-Mistral-7B model that has been further fine-tuned with Direct Preference Optimization (DPO) using the mlabonne/chatml_dpo_pairs dataset. It surpasses the original model on most benchmarks (see results).

It is directly inspired by the RLHF process described by Intel/neural-chat-7b-v3-1's authors to improve performance. I used the same dataset and reformatted it to apply the ChatML template.

The code to train this model is available on Google Colab and GitHub. It required an A100 GPU for about an hour.

Quantized models

Results

Update: NeuralHermes-2.5 became the best Hermes-based model on the Open LLM leaderboard and one of the very best 7b models. 🎉

image/png

Teknium (author of OpenHermes-2.5-Mistral-7B) benchmarked the model (see his tweet).

Results are improved on every benchmark: AGIEval (from 43.07% to 43.62%), GPT4All (from 73.12% to 73.25%), and TruthfulQA.

AGIEval

GPT4All

TruthfulQA

You can check the Weights & Biases project here.

Usage

You can run this model using LM Studio or any other frontend.

You can also run this model using the following code:

import transformers
from transformers import AutoTokenizer

# Format prompt
message = [
    {"role": "system", "content": "You are a helpful assistant chatbot."},
    {"role": "user", "content": "What is a Large Language Model?"}
]
tokenizer = AutoTokenizer.from_pretrained(new_model)
prompt = tokenizer.apply_chat_template(message, add_generation_prompt=True, tokenize=False)

# Create pipeline
pipeline = transformers.pipeline(
    "text-generation",
    model=new_model,
    tokenizer=tokenizer
)

# Generate text
sequences = pipeline(
    prompt,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    num_return_sequences=1,
    max_length=200,
)
print(sequences[0]['generated_text'])

Training hyperparameters

LoRA:

  • r=16
  • lora_alpha=16
  • lora_dropout=0.05
  • bias="none"
  • task_type="CAUSAL_LM"
  • target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']

Training arguments:

  • per_device_train_batch_size=4
  • gradient_accumulation_steps=4
  • gradient_checkpointing=True
  • learning_rate=5e-5
  • lr_scheduler_type="cosine"
  • max_steps=200
  • optim="paged_adamw_32bit"
  • warmup_steps=100

DPOTrainer:

  • beta=0.1
  • max_prompt_length=1024
  • max_length=1536