Hugging Face's logo

pszemraj
/
mGPT-Peter-2E

Text Generation
Transformers
PyTorch
Safetensors
gpt2
multilingual
PyTorch
Transformers
gpt3
Deepspeed
Megatron
mGPT
text-generation-inference
Inference Endpoints
Model card Files Community
1
mGPT-Peter-2E / README.md
pszemraj's picture
pszemraj
Update README.md
7663e8d
preview code
|
raw
history blame
4.41 kB
metadata
license: apache-2.0
pipeline_tag: text-generation
tags:
  - multilingual
  - PyTorch
  - Transformers
  - gpt3
  - gpt2
  - Deepspeed
  - Megatron
  - mGPT
datasets:
  - mc4
  - Wikipedia
widget:
  - text: >-
      Ich weiß, dass du müde bist, aber können wir heute Abend noch einen
      Spaziergang machen? peter szemraj: ich
    example_title: walk - Deutsch
  - text: 'peter szemraj: 我喜欢穿很酷的衣服'
    example_title: fashion - Chinese
  - text: 'Wat zei je over mijn moeder? peter szemraj: ik'
    example_title: 🚎 - Dutch
  - text: >-
      Zagadka: Człowiekowi, który przebywał na dworze w deszczu bez parasola czy
      kapelusza, nie zmoczył się ani jeden włos na głowie. Dlaczego? peter
      szemraj: czy to
    example_title: brain teaser - Polish
  - text: >-
      Minha amiga diz que conhece todas as línguas, mas não fala nenhuma
      delas... o que há de errado com ela? peter szemraj: eu
    example_title: language - Portuguese
  - text: 'se potesse vivere ovunque, dove sarebbe? peter szemraj: io'
    example_title: dream living place - Italian
  - text: 'Can you take me for dinner somewhere nice this time? peter szemraj:'
    example_title: dinner
  - text: 'What really makes you angry? peter szemraj:'
    example_title: pet peeve
  - text: >-
      Jak nazwać aligatora, który właśnie przeszedł operację usunięcia lewego
      ramienia?peter szemraj: ja
    example_title: alligator - Polish
  - text: >-
      Warum sind Transformers für die Sprachmodellierung wichtig?  peter
      szemraj: es ist
    example_title: Transformers - German
  - text: >-
      как написать хорошие подсказки для языковых моделей? peter szemraj:
      сначала вам нужно
    example_title: prompt tutorial - Russian
  - text: >-
      Pewien mężczyzna wpycha swój samochód do hotelu i mówi właścicielowi, że
      jest bankrutem. Dlaczego? peter szemraj: może
    example_title: brain teaser - Polish 2
  - text: >-
      Zagadka: Mówię bez ust i słyszę bez uszu. Nie mam ciała, ale ożywiam się
      wraz z wiatrem. Czym jestem? peter szemraj: czy to
    example_title: brain teaser - Polish 3
  - text: 'Què t''agrada fer per divertir-te? peter szemraj: m''agrada'
    example_title: hobbies - Catalan
  - text: '为什么你总是那么累？peter szemraj: 呃，我想'
    example_title: tired - Chinese
inference:
  parameters:
    min_length: 2
    max_length: 64
    do_sample: true
    top_k: 10
    top_p: 0.9
    temperature: 0.65
    repetition_penalty: 3.5
    no_repeat_ngram_size: 3
    length_penalty: 0.4
    pad_token: 1

mGPT: fine-tune on message data - 2E

  • This model is a fine-tuned version of sberbank-ai/mGPT on 80k messages. This builds on the minimum-working-example checkpoint here.
  • 2E = 2 epochs

Model description

  • testing if fine-tuned personality data bleeds over to other languages without being trained in them explicitly

Interesting findings thus far:

  • Passing a generic word after the <name-identifier> that is in a non-English language helps ensure the model responds in the question language (see: any example).
  • Model generations (in general) remain semantically consistent, even if the generations switch from <language>to English in the middle of the generated text. This demonstrates some sort of "universal concept understanding"

Usage in python

Install the transformers library if you don't have it:

pip install -U transformers

load the model into a pipeline object:

from transformers import pipeline
import torch
device = 'cuda' if torch.cuda.is_available() else 'cpu'
my_chatbot = pipeline('text-generation', 
                      'pszemraj/mGPT-Peter-2E',
                      device=0 if device == 'cuda' else -1,
                    )

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-05
  • train_batch_size: 4
  • eval_batch_size: 4
  • seed: 42
  • distributed_type: multi-GPU
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine_with_restarts
  • lr_scheduler_warmup_ratio: 0.05
  • num_epochs: 1 (in addition to all training on prior checkpoints)

Framework versions

  • Transformers 4.18.0
  • Pytorch 1.11.0+cu113
  • Datasets 2.1.0
  • Tokenizers 0.12.1