|
--- |
|
license: |
|
- creativeml-openrail-m |
|
language: |
|
- en |
|
tags: |
|
- generated_from_trainer |
|
- text generation |
|
- pytorch |
|
- casual-lm |
|
metrics: |
|
- accuracy |
|
model-index: |
|
- name: openchatgpt-neox-r1 |
|
results: [] |
|
--- |
|
|
|
# openchatgpt-neox-r1 |
|
|
|
This model is a fine-tuned version of [EleutherAI/pythia-125m-deduped](https://huggingface.co/EleutherAI/pythia-125m-deduped) on the openchatgpt safe-r1 dataset. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 1.3585 |
|
- Accuracy: 0.9169 |
|
|
|
## Model description |
|
|
|
Finetune based on the inner workings of ChatGPT. I won't elaborate on that. You must have a faint idea of how prompt is made for it to spit anything that's not garbled mess. |
|
|
|
This is effectively a schizophrenic idea that met the light of day. Practically a collab of 3 students in a virtual shed. |
|
|
|
BTW, Pythia is so much better omg. |
|
|
|
## Intended uses & limitations |
|
|
|
Intended uses & limitations fall in line with OpenAI's. Dataset used consists of safe texts (i.e. not highly sexual/erotica type stuff). NSFW version of the dataset is not planned to exist at the moment. |
|
|
|
Keep in mind that this is a 125m version of GPT-NeoX (Pythia). My 1050Ti Mobile couldn't even handle that without gradient thingmabobs, 8BitAdam was also used. If anyone knows how to effectively finetune larger models on free colabs - feel free to let me know. Pile tokenizer also has one downside compared to native GPT-2/3 - `Assistant` is not 1 token, but 2. |
|
|
|
## Training and evaluation data |
|
|
|
Data was split in ratio of 95%/5%. Preproccess included removing mentions of OpenAI wherever it was not deemed appropriete (GPT-2 has one of the appropriete mentions). Whole dataset consists of just shy off 3k input-output pairs. One input has multiple outputs (read as: one message has multiple variants of an answer). <<<1% (3 total) are curated lines (i.e. a huge mistake was spotted that needed corrections). At least 3 lines (<<<1% of line count, but more of byte count) are broken. |
|
|
|
Heavy bias on IT. |
|
|
|
## Training procedure |
|
|
|
Input and output were straight up concatenated due to the nature of how ChatGPT works. |
|
|
|
This time dataset was batched into groups of 2048 tokens. Meaning i got 628/31 groups for training/eval. Maybe that's what made the difference. EOS was also being used after the final separator. |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 5e-05 |
|
- train_batch_size: 1 |
|
- eval_batch_size: 1 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 2 |
|
- total_train_batch_size: 2 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- num_epochs: 3.0 |
|
- mixed_precision_training: Native AMP |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Accuracy | |
|
|:-------------:|:-----:|:----:|:---------------:|:--------:| |
|
| 1.1311 | 1.0 | 1377 | 1.3116 | 0.9127 | |
|
| 0.6691 | 2.0 | 2754 | 1.2978 | 0.9160 | |
|
| 0.3463 | 3.0 | 4131 | 1.3585 | 0.9169 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.25.1 |
|
- Pytorch 1.13.1+cu116 |
|
- Datasets 2.8.0 |
|
- Tokenizers 0.13.2 |
|
|
