|
--- |
|
license: openrail |
|
datasets: |
|
- Locutusque/ColumnedChatCombined |
|
language: |
|
- en |
|
metrics: |
|
- bleu |
|
- perplexity |
|
- loss |
|
- reward |
|
- penalty |
|
widget: |
|
- text: "<|USER|> Hello! <|ASSISTANT|> " |
|
pipeline_tag: text-generation |
|
--- |
|
# Model Card |
|
## Model Details |
|
- Model Name: gpt2-conversational-or-qa (prototype) |
|
- Model Type: Language Modeling |
|
- Task: Generating Conversational Responses |
|
- Description: This model is trained on a dataset of conversations between a user and an AI assistant, with the goal of generating a coherent and relevant response to the user's input. It uses the GPT-2 architecture, a state-of-the-art transformer-based language model that is capable of generating high-quality text with a wide range of styles and tones. The model is fine-tuned on the conversational data using maximum likelihood estimation, and is evaluated based on its ability to generate responses that are both grammatically correct and semantically relevant to the user's input. |
|
## Intended Use |
|
This model is intended to be used for generating conversational responses in a variety of contexts, such as chatbots, virtual assistants, and customer service applications. It is designed to provide natural and engaging responses to user input, with a focus on maintaining a consistent tone and style throughout the conversation. The model is suitable for use in both text-based and voice-based interfaces, and can be easily integrated into existing applications using the PyTorch and Transformers frameworks. |
|
|
|
## Training Data |
|
The model is trained on a large dataset of conversational data, consisting of interactions between users and an AI assistant. The data is preprocessed to remove any sensitive information and is formatted in a way that is suitable for training a language model. The training data is split into a training set and a validation set, with the training set used to update the model parameters and the validation set used to evaluate the model performance. The model was trained on 245,000 examples over 1,225,000 steps, it achieved decent metrics. |
|
This model outperformed the base GPT-2 model significantly on a new conversational dataset during a fine-tuning session. Here is a side-by-side comparison of the two models during the first steps of training |
|
```python |
|
# Base GPT-2 |
|
""" |
|
Epoch 1/5, Batch 1/10000: Loss - 64.9255, Reward - 260.0000, Penalty - 624.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 2/10000: Loss - 57.4635, Reward - 303.0000, Penalty - 870.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 3/10000: Loss - 67.8061, Reward - 295.0000, Penalty - 908.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 4/10000: Loss - 59.6118, Reward - 800.0000, Penalty - 740.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 5/10000: Loss - 67.4855, Reward - 402.0000, Penalty - 806.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 6/10000: Loss - 29.3718, Reward - 937.0000, Penalty - 760.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 7/10000: Loss - 79.0709, Reward - 390.0000, Penalty - 1114.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 8/10000: Loss - 61.4583, Reward - 385.0000, Penalty - 760.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 9/10000: Loss - 56.3084, Reward - 741.0000, Penalty - 560.0000, BLEU - 3.5500 |
|
Epoch 1/5, Batch 10/10000: Loss - 80.0192, Reward - 838.0000, Penalty - 1424.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 11/10000: Loss - 51.8236, Reward - 228.0000, Penalty - 812.0000, BLEU - 0.0001 |
|
Epoch 1/5, Batch 12/10000: Loss - 71.4071, Reward - 541.0000, Penalty - 982.0000, BLEU - 0.0000 |
|
Epoch 1/5, Batch 13/10000: Loss - 33.3624, Reward - 910.0000, Penalty - 1002.0000, BLEU - 0.0027 |
|
Epoch 1/5, Batch 14/10000: Loss - 55.9721, Reward - 808.0000, Penalty - 798.0000, BLEU - 0.0005 |
|
Epoch 1/5, Batch 15/10000: Loss - 67.0336, Reward - 517.0000, Penalty - 764.0000, BLEU - 0.0000 |
|
""" |
|
# Conversational GPT-2 |
|
""" |
|
Epoch 1/5, Batch 1/10000: Loss - 6.1980, Reward - 887.0000, Penalty - 1500.0000, BLEU - 0.0648 |
|
Epoch 1/5, Batch 2/10000: Loss - 4.5750, Reward - 245.0000, Penalty - 1618.0000, BLEU - 0.0008 |
|
Epoch 1/5, Batch 3/10000: Loss - 5.1264, Reward - 600.0000, Penalty - 642.0000, BLEU - 5.7981 |
|
Epoch 1/5, Batch 4/10000: Loss - 0.2995, Reward - 1020.0000, Penalty - 74.0000, BLEU - 13.8469 |
|
Epoch 1/5, Batch 5/10000: Loss - 7.9377, Reward - 203.0000, Penalty - 1700.0000, BLEU - 0.3218 |
|
Epoch 1/5, Batch 6/10000: Loss - 5.0522, Reward - 1020.0000, Penalty - 2034.0000, BLEU - 0.1946 |
|
Epoch 1/5, Batch 7/10000: Loss - 2.0585, Reward - 925.0000, Penalty - 526.0000, BLEU - 16.1298 |
|
Epoch 1/5, Batch 8/10000: Loss - 5.9736, Reward - 1009.0000, Penalty - 1844.0000, BLEU - 0.0085 |
|
Epoch 1/5, Batch 9/10000: Loss - 6.0867, Reward - 245.0000, Penalty - 1690.0000, BLEU - 1.9342 |
|
Epoch 1/5, Batch 10/10000: Loss - 7.8497, Reward - 155.0000, Penalty - 1780.0000, BLEU - 0.0115 |
|
Epoch 1/5, Batch 11/10000: Loss - 3.8887, Reward - 1012.0000, Penalty - 2010.0000, BLEU - 0.6957 |
|
Epoch 1/5, Batch 12/10000: Loss - 6.6133, Reward - 216.0000, Penalty - 1638.0000, BLEU - 1.7853 |
|
Epoch 1/5, Batch 13/10000: Loss - 1.3319, Reward - 945.0000, Penalty - 374.0000, BLEU - 0.0075 |
|
Epoch 1/5, Batch 14/10000: Loss - 2.6296, Reward - 956.0000, Penalty - 414.0000, BLEU - 3.2207 |
|
Epoch 1/5, Batch 15/10000: Loss - 6.8827, Reward - 1013.0000, Penalty - 1970.0000, BLEU - 3.7418 |
|
""" |
|
``` |
|
## Model Architecture |
|
The model architecture used in this model is GPT-2, a transformer-based language model that is capable of generating high-quality text with a wide range of styles and tones. The GPT-2 architecture consists of a multi-layered transformer encoder-decoder, with self-attention mechanisms that allow the model to capture long-term dependencies and generate coherent text. |
|
|
|
## Evaluation Metrics |
|
The model is evaluated based on several metrics, including loss, reward, penalty, BLEU score, and perplexity. The loss metric is calculated during training and reflects the difference between the predicted output and the actual output. The reward metric is based on the number of correct words generated by the model, while the penalty metric penalizes the model for repeating words consecutively. The BLEU score measures the similarity between the generated text and the ground truth text, while the perplexity metric measures how well the model is able to predict the next word in a sequence. During validation, the model achieved the following metrics: |
|
|
|
- BLEU Score: 9 |
|
- Perplexity: 19 |
|
- Loss: 1.7 |
|
|
|
Although these metrics seem mediocre, it's actually better because that way the model is able to make open-ended responses, but is still coherent to the user's input. |
|
|
|
## Limitations and Bias |
|
This model is not suitable for all use cases due to its limited training time on a weak computer. As a result, it may produce irrelevant or nonsensical responses. Additionally, it has not been fine-tuned to remember the chat history, is unable to provide follow-up responses, and it does not know the answer to many questions (it was only fine-tuned to respond in a conversational way). For optimal performance, we recommend using a GPU with at least 4GB of VRAM and downloading the model manually instead of using the Transformers library or deploying it on the Interface API. Here's how you should deploy the model: |
|
|
|
```python |
|
import torch |
|
from transformers import GPT2Tokenizer, GPT2LMHeadModel |
|
|
|
tokenizer = GPT2Tokenizer.from_pretrained('gpt2') |
|
model = GPT2LMHeadModel.from_pretrained('gpt2') |
|
tokenizer.add_special_tokens({'pad_token': '[PAD]'}) |
|
tokenizer.add_special_tokens({'eos_token': '<|End|>'}) |
|
special_tokens = { |
|
"additional_special_tokens": ["<|USER|>", "<|SYSTEM|>", "<|ASSISTANT|>"] |
|
} |
|
tokenizer.add_special_tokens(special_tokens) |
|
model.resize_token_embeddings(len(tokenizer)) |
|
model.load_state_dict(torch.load("path/to/model")) |
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
model.to(device) |
|
def generate_text(model, tokenizer, prompt, max_length=1024): |
|
prompt = f'<|USER|> {prompt} <|ASSISTANT|> ' |
|
input_ids = tokenizer.encode(prompt, add_special_tokens=True, return_tensors="pt").to(device) |
|
attention_mask = torch.ones_like(input_ids).to(device) |
|
output = model.generate(input_ids, |
|
max_length=max_length, |
|
do_sample=True, |
|
top_k=35, |
|
top_p=0.80, |
|
pad_token_id=tokenizer.pad_token_id, |
|
eos_token_id=tokenizer.eos_token_id, |
|
attention_mask=attention_mask) |
|
output_ids = tokenizer.decode(output[0], skip_special_tokens=False) |
|
assistant_token_index = output_ids.index('<|ASSISTANT|>') + len('<|ASSISTANT|>') |
|
next_token_index = output_ids.find('<|', assistant_token_index) |
|
output_ids = output_ids[assistant_token_index:next_token_index] |
|
return output_ids |
|
# Loop to interact with the model |
|
while True: |
|
prompt = input("Enter a prompt (or 'q' to quit): ") |
|
if prompt == "q": |
|
break |
|
output_text = generate_text(model, tokenizer, prompt) |
|
print(output_text) |
|
``` |
|
## Deploying and training the model |
|
The model has been fine-tuned on a specific input format that goes like this ```"<|USER|> {user prompt} <|ASSISTANT|> {model prediction} <|End|>".``` For the best performance from the model the input text should be as follows ```<|USER|> {user prompt} <|ASSISTANT|> ``` and the target/label should be as follows ```<|USER|> {user prompt} <|ASSISTANT|> {dataset output} <|End|>``` |
