language:
- en
- vi
license: mit
library_name: transformers
tags:
- ghost
- awq
pipeline_tag: text-generation
model-index:
- name: ghost-7b-v0.9.1
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: 55.38
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=lamhieu/ghost-7b-v0.9.1
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: 77.03
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=lamhieu/ghost-7b-v0.9.1
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: 54.78
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=lamhieu/ghost-7b-v0.9.1
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: 43.96
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=lamhieu/ghost-7b-v0.9.1
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: 72.53
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=lamhieu/ghost-7b-v0.9.1
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: 26.91
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=lamhieu/ghost-7b-v0.9.1
name: Open LLM Leaderboard
widget:
- text: How many helicopters can a human eat in one sitting
output:
text: >-
Ahoy, me matey! A human can eat approximately one helicopter in one
sitting, but only if they're a giant sea monster with a stomach the size
of a small country. 🤢🤢 So, it's not advisable to try this, pirate!
🏰🛢️
model_creator: Hieu Lam
model_name: Ghost 7B v0.9.1
model_type: mistral
prompt_template: <|system|>\n</s>\n<|user|>\n{prompt}</s><|assistant|>\n
Model Card for Model ID
Ghost 7B Alpha, flying, v0.9.1, AWQ.
Come on, create yourself an AI assistant, according to your wishes!
In your language, maybe Vietnamese.
Or, English.
Let the assistant become an expert, and more.
The challenge of the model's ability to understand the language.
Challenge the model's reasoning ability, in Vietnamese language.
In case of using Vietnamese language, it lacks accents, abbreviations or uses slang.
📚 Model Details
Model Description
A version to consider comprehension in generating languages other than the original language being initially trained, here is the Vietnamese language. A brief summary of the effectiveness of the Mistral 7B model for training with a new language is excellent and low cost.
I have started training the Ghost 7B v0.9.0 model again, with a smaller amount of data, it is estimated to be only about 150MB. In that data, about 70% is Vietnamese, the rest is almost English. The approach here uses QLora for training then merges them. Also, I am very thankful to Unsloth for their features.
About AWQ and AutoAWQ
The AWQ algorithm for 4-bit quantization with a 2x speedup during inference.
AutoAWQ is an easy-to-use package for 4-bit quantized models. AutoAWQ speeds up models by 3x and reduces memory requirements by 3x compared to FP16. AutoAWQ implements the Activation-aware Weight Quantization (AWQ) algorithm for quantizing LLMs. AutoAWQ was created and improved upon from the original work from MIT.
@article{lin2023awq,
title={AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration},
author={Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and Dang, Xingyu and Han, Song},
journal={arXiv},
year={2023}
}
⛹️♂️ Uses
Online using Google Colab
To make it easier to play around with the model, I created a notebook in Google Colab so you can start experimenting.
Directly
For direct use, you can easily get started with the following steps.
Firstly, you need to install transformers via the command below with
pip.pip install -U transformersRight now, you can start using the model directly.
import torch from transformers import ( AutoModelForCausalLM, AutoTokenizer, ) base_model = "lamhieu/ghost-7b-v0.9.1" model = AutoModelForCausalLM.from_pretrained( base_model, torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto", ) tokenizer = AutoTokenizer.from_pretrained(base_model) messages = [ {"role": "system", "content": "You are a friendly chatbot who always responds in the style of a pirate"}, {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) tokenized = tokenizer(prompt, return_tensors="pt", add_special_tokens=False) outputs = model.generate(**tokenized, max_new_tokens=512) results = tokenizer.batch_decode(outputs)[0] print(results)Additionally, you can also use a model with 4bit quantization to reduce the required resources at least. You can start with the code below.
import torch from transformers import ( AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, ) base_model = "lamhieu/ghost-7b-v0.9.1" bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=False, ) model = AutoModelForCausalLM.from_pretrained( base_model, quantization_config=bnb_config, trust_remote_code=True, device_map="auto", ) tokenizer = AutoTokenizer.from_pretrained(base_model) messages = [ {"role": "system", "content": "You are a friendly chatbot who always responds in the style of a pirate"}, {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) tokenized = tokenizer(prompt, return_tensors="pt", add_special_tokens=False) outputs = model.generate(**tokenized, max_new_tokens=512) results = tokenizer.batch_decode(outputs)[0] print(results)
Summary
Although the amount of training data is small, it is "great". You don't need to worry too much that it won't be able to meet some of your requirements. Instead, try experimenting with the model of what you want. One more thing, use it like you would ChatGPT, I've purposely tweaked it to be able to replace my app (for some tasks, and it does a good job). It's okay with both Vietnamese and English languages. It would be great to hear feedback about the experience, feel free to leave information in the discussion section.
Setting up the system prompt will have a great impact on the performance and quality of the content generated by the model. Keep this in mind to always ensure the model is used for your intended purpose, the goal is to achieve good results but. It's best to always set system, you can still leave it empty if you always want to set it.
🥇 Evaluation
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 55.10 |
| AI2 Reasoning Challenge (25-Shot) | 55.38 |
| HellaSwag (10-Shot) | 77.03 |
| MMLU (5-Shot) | 54.78 |
| TruthfulQA (0-shot) | 43.96 |
| Winogrande (5-shot) | 72.53 |
| GSM8k (5-shot) | 26.91 |
VMLU
A Vietnamese Multitask Language Understanding Benchmark Suite for Large Language Models.
With the score achieved, the model can rank 3rd in VMLU's "Leaderboard of fine-tuned models" list, as of the date of evaluation.
Details
{
"humanity": {
"administrative_law": 52.22,
"business_law": 40.22,
"civil_law": 46.11,
"criminal_law": 49.08,
"economic_law": 39.75,
"education_law": 42.17,
"elementary_history": 55.37,
"high_school_history": 36.67,
"high_school_literature": 37.78,
"history_of_world_civilization": 46.67,
"idealogical_and_moral_cultivation": 50,
"introduction_to_laws": 45.24,
"vietnamese_language_and_literature": 34.48,
"total": 43.3,
"revolutionary_policy_of_the_vietnamese_commununist_part": 51.11,
"introduction_to_vietnam_culture": 30.56,
"logic": 27.01,
"middle_school_history": 44.44,
"middle_school_literature": 50.57
},
"stem": {
"total": 34.73,
"applied_informatics": 50.56,
"computer_architecture": 33.89,
"computer_network": 43.02,
"discrete_mathematics": 31.52,
"electrical_engineering": 30.68,
"elementary_mathematics": 30,
"elementary_science": 58.89,
"high_school_biology": 38.33,
"high_school_chemistry": 28.89,
"high_school_mathematics": 26.35,
"high_school_physics": 29.44,
"introduction_to_chemistry": 27.37,
"introduction_to_physics": 31.79,
"introduction_to_programming": 36.31,
"metrology_engineer": 31.21,
"middle_school_biology": 46.47,
"middle_school_chemistry": 30.56,
"middle_school_mathematics": 30.56,
"middle_school_physics": 30,
"operating_system": 40.56,
"statistics_and_probability": 22.99
},
"total": 39.58,
"other": {
"accountant": 31.55,
"civil_servant": 42.11,
"clinical_pharmacology": 33.89,
"driving_license_certificate": 59.06,
"environmental_engineering": 28.07,
"internal_basic_medicine": 39.77,
"preschool_pedagogy": 46.08,
"tax_accountant": 22.41,
"tax_civil_servant": 47.95,
"total": 38.99
},
"social_science": {
"business_administration": 41.38,
"high_school_civil_education": 45,
"high_school_geography": 34.57,
"ho_chi_minh_ideology": 48.04,
"macroeconomics": 31.11,
"microeconomics": 37.22,
"middle_school_civil_education": 66.29,
"middle_school_geography": 48.3,
"principles_of_marxism_and_leninism": 30,
"sociology": 53.93,
"total": 43.58
}
}
📜 More Information
Note, this is a personal research project with a limited budget, so the model only stops at the evaluation level with the developed approach. Apart from that, I think I can definitely build a model with better quality in terms of language and other performance using this approach.
Thanks for the support
Model trained with Unsloth, many thanks.
📨 Model Card Contact
Lam Hieu (lamhieu.vk@gmail.com)