274 lines
8.5 KiB
Markdown
274 lines
8.5 KiB
Markdown
---
|
||
language:
|
||
- ja
|
||
- en
|
||
- zh
|
||
license: apache-2.0
|
||
model-index:
|
||
- name: laser-polyglot-4x7b
|
||
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: 64.16
|
||
name: normalized accuracy
|
||
source:
|
||
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=macadeliccc/laser-polyglot-4x7b
|
||
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: 84.98
|
||
name: normalized accuracy
|
||
source:
|
||
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=macadeliccc/laser-polyglot-4x7b
|
||
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: 63.88
|
||
name: accuracy
|
||
source:
|
||
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=macadeliccc/laser-polyglot-4x7b
|
||
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: 55.47
|
||
source:
|
||
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=macadeliccc/laser-polyglot-4x7b
|
||
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: 77.82
|
||
name: accuracy
|
||
source:
|
||
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=macadeliccc/laser-polyglot-4x7b
|
||
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: 48.45
|
||
name: accuracy
|
||
source:
|
||
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=macadeliccc/laser-polyglot-4x7b
|
||
name: Open LLM Leaderboard
|
||
---
|
||
# Polyglot-4x7b-24b
|
||
|
||
|
||
|
||
Polyglot-4x7b is a Mixture of Experts approach to a multilingual model.
|
||
|
||
This project is an experiment to see if each expert can be of a different language. The answer is yes.
|
||
|
||
The model is a merge of models that are capable of Chinese and Japanese output.
|
||
|
||
+ teknium/OpenHermes-2.5-Mistral-7B
|
||
+ oshizo/japanese-e5-mistral-7b_slerp
|
||
+ cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser
|
||
+ s3nh/Mistral-7B-Evol-Instruct-Chinese
|
||
|
||
TODO:
|
||
1. [] polyglot tokenizer
|
||
|
||
## Other polyglot models
|
||
|
||
+ [macadeliccc/Polyglot-8x7b-v0.1](https://huggingface.co/macadeliccc/Polyglot-8x7b-v0.1) (adds 3 more languages)
|
||
# Code Example
|
||
|
||
Inference [Colab](https://colab.research.google.com/drive/1tYSb63IKZDsiQ5BIJU8Oc92phxugAmB3?usp=sharing)
|
||
Live demo available on [Spaces](https://huggingface.co/spaces/macadeliccc/polyglot-4x7b-chat?logs=build)
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
def generate_response(prompt):
|
||
"""
|
||
Generate a response from the model based on the input prompt.
|
||
|
||
Args:
|
||
prompt (str): Prompt for the model.
|
||
|
||
Returns:
|
||
str: The generated response from the model.
|
||
"""
|
||
# Tokenize the input prompt
|
||
inputs = tokenizer(prompt, return_tensors="pt")
|
||
|
||
# Generate output tokens
|
||
outputs = model.generate(**inputs, max_new_tokens=256, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id)
|
||
|
||
# Decode the generated tokens to a string
|
||
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||
|
||
return response
|
||
|
||
# Load the model and tokenizer
|
||
model_id = "macadeliccc/laser-polyglot-4x7b"
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True)
|
||
|
||
# Example prompts in different languages
|
||
english_prompt = "Write a quicksort algorithm in python"
|
||
chinese_prompt = "用Python写一个快速排序算法"
|
||
japanese_prompt = "Pythonでクイックソートアルゴリズムを書いてください"
|
||
|
||
# Generate and print responses for each language
|
||
print("English Response:")
|
||
print(generate_response(english_prompt), "\n")
|
||
|
||
print("Chinese Response:")
|
||
print(generate_response(chinese_prompt), "\n")
|
||
|
||
print("Japanese Response:")
|
||
print(generate_response(japanese_prompt), "\n")
|
||
|
||
```
|
||
|
||
## Example Output
|
||
English:
|
||
|
||
Write a quicksort algorithm in python.
|
||
|
||
```python
|
||
def quicksort(arr):
|
||
if len(arr) <= 1:
|
||
return arr
|
||
else:
|
||
pivot = arr[0]
|
||
less = [i for i in arr[1:] if i <= pivot]
|
||
greater = [i for i in arr[1:] if i > pivot]
|
||
return quicksort(less) + [pivot] + quicksort(greater)
|
||
|
||
arr = [5, 2, 9, 1, 5, 7, 4, 8, 6, 3]
|
||
print(quicksort(arr))
|
||
```
|
||
|
||
This is a simple implementation of the quicksort algorithm in python. The function `quicksort` takes an array as input and returns a sorted array. The algorithm works by selecting a pivot element from the array and partitioning the other elements into two sub-arrays, according to whether they are less than or greater than the pivot. The process is then repeated recursively on the sub-arrays until the entire array is sorted.
|
||
|
||
Chinese Response:
|
||
|
||
用Python写一个快速排序算法
|
||
|
||
```python
|
||
def quick_sort(arr):
|
||
if len(arr) <= 1:
|
||
return arr
|
||
else:
|
||
pivot = arr[0]
|
||
less = [i for i in arr[1:] if i <= pivot]
|
||
greater = [i for i in arr[1:] if i > pivot]
|
||
return quick_sort(less) + [pivot] + quick_sort(greater)
|
||
|
||
arr = [3, 5, 2, 1, 4, 6, 8, 7]
|
||
print(quick_sort(arr))
|
||
```
|
||
这个程序的时间复杂度为O(nlogn)，空间复杂度为O(n)。
|
||
|
||
Japanese Response:
|
||
|
||
Pythonでクイックソートアルゴリズムを書いてください。
|
||
|
||
```python
|
||
def quicksort(arr):
|
||
if len(arr) <= 1:
|
||
return arr
|
||
pivot = arr[0]
|
||
left = [x for x in arr[1:] if x < pivot]
|
||
right = [x for x in arr[1:] if x >= pivot]
|
||
return quicksort(left) + [pivot] + quicksort(right)
|
||
|
||
print(quicksort([3,6,8,10,1,5,9,2,4,7]))
|
||
```
|
||
|
||
このコードはクイックソートアルゴリズムを実装しています。クイックソートは一種の分割と conquers アルゴリズムで、配列を分割し、それぞれの部分配列を再帰的にソートします。
|
||
|
||
この実装では、配列の最初の要素をピボットとして使用します。そして、配列を2つの
|
||
|
||
|
||
|
||
# Evaluations
|
||
|
||
| Tasks |Version|Filter|n-shot| Metric |Value | |Stderr|
|
||
|-------------|-------|------|-----:|--------|-----:|---|-----:|
|
||
|arc_challenge|Yaml |none | 0|acc |0.5495|± |0.0145|
|
||
| | |none | 0|acc_norm|0.5794|± |0.0144|
|
||
|arc_easy |Yaml |none | 0|acc |0.8304|± |0.0077|
|
||
| | |none | 0|acc_norm|0.8068|± |0.0081|
|
||
|boolq |Yaml |none | 0|acc |0.8749|± |0.0058|
|
||
|hellaswag |Yaml |none | 0|acc |0.6276|± |0.0048|
|
||
| | |none | 0|acc_norm|0.8157|± |0.0039|
|
||
|openbookqa |Yaml |none | 0|acc |0.3180|± |0.0208|
|
||
| | |none | 0|acc_norm|0.4460|± |0.0223|
|
||
|piqa |Yaml |none | 0|acc |0.8139|± |0.0091|
|
||
| | |none | 0|acc_norm|0.8237|± |0.0089|
|
||
|winogrande |Yaml |none | 0|acc |0.7419|± |0.0123|
|
||
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
||
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_macadeliccc__laser-polyglot-4x7b)
|
||
|
||
| Metric |Value|
|
||
|---------------------------------|----:|
|
||
|Avg. |65.79|
|
||
|AI2 Reasoning Challenge (25-Shot)|64.16|
|
||
|HellaSwag (10-Shot) |84.98|
|
||
|MMLU (5-Shot) |63.88|
|
||
|TruthfulQA (0-shot) |55.47|
|
||
|Winogrande (5-shot) |77.82|
|
||
|GSM8k (5-shot) |48.45|
|
||
|