98 lines
2.8 KiB
Markdown
98 lines
2.8 KiB
Markdown
|
---
|
|||
|
|
license: mit
|
|||
|
|
base_model: vicgalle/gpt2-open-instruct-v1
|
|||
|
|
tags:
|
|||
|
|
- generated_from_trainer
|
|||
|
|
- Transformers
|
|||
|
|
- GPT2
|
|||
|
|
model-index:
|
|||
|
|
- name: hh-rlhf
|
|||
|
|
results: []
|
|||
|
|
datasets:
|
|||
|
|
- Anthropic/hh-rlhf
|
|||
|
|
- hakurei/open-instruct-v1
|
|||
|
|
tokenizers:
|
|||
|
|
- GPT2Tokenizer
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
library_name: transformers
|
|||
|
|
metrics:
|
|||
|
|
- bleu
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
|||
|
|
should probably proofread and complete it, then remove this comment. -->
|
|||
|
|
|
|||
|
|
# hh-rlhf
|
|||
|
|
|
|||
|
|
This model is a fine-tuned version of [vicgalle/gpt2-open-instruct-v1](https://huggingface.co/vicgalle/gpt2-open-instruct-v1) on an subset (15k) of the Anthropic/hh-rlhf dataset.
|
|||
|
|
It achieves the following results on the evaluation set:
|
|||
|
|
- Loss: 2.1534
|
|||
|
|
|
|||
|
|
This model responds to the 'Human:' or 'Assistant:' prompt pretty well in conversation situations.
|
|||
|
|
|
|||
|
|
The shorter responses are better suited. Keep generation length to a reasonable subset. Left to its own devices it will have some pretty esoteric responses.
|
|||
|
|
|
|||
|
|
These include fairly uncensored remarks and at times violent outbursts. Especially if asking questions.
|
|||
|
|
|
|||
|
|
Needs vetting for other textual uses.
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
Human: Insane clown posse says...
|
|||
|
|
|
|||
|
|
Human: Should we look for a woman?
|
|||
|
|
|
|||
|
|
Assistant: It’s okay if you’re having a tough time finding what you are looking for. It’s a common question people might come up with for an argument or misunderstanding. What are you looking for, and what kind of woman would you have?
|
|||
|
|
|
|||
|
|
Human: Are you trying to find someone to argue
|
|||
|
|
```
|
|||
|
|
## Model description
|
|||
|
|
|
|||
|
|
GPT2 open instruct was trained on the open-instruct dataset fully. The reimagines one LM head as a partial rhlf adapter, with subtle reinforcements.
|
|||
|
|
## Intended uses & limitations
|
|||
|
|
|
|||
|
|
Intended to study the intersection of instruct models and prompting that focuses on subtle exchanges of prompting. This probably needs to be refined substantially at this point.
|
|||
|
|
|
|||
|
|
## Training and evaluation data
|
|||
|
|
```python
|
|||
|
|
Train dataset size: 15000
|
|||
|
|
Test dataset size: 500
|
|||
|
|
Dataset({
|
|||
|
|
features: ['chosen', 'rejected'],
|
|||
|
|
num_rows: 15000
|
|||
|
|
})
|
|||
|
|
Dataset({
|
|||
|
|
features: ['chosen', 'rejected'],
|
|||
|
|
num_rows: 500
|
|||
|
|
})
|
|||
|
|
```
|
|||
|
|
## Training procedure
|
|||
|
|
|
|||
|
|
### Training hyperparameters
|
|||
|
|
|
|||
|
|
The following hyperparameters were used during training:
|
|||
|
|
- learning_rate: 0.0005
|
|||
|
|
- train_batch_size: 2
|
|||
|
|
- eval_batch_size: 1
|
|||
|
|
- seed: 42
|
|||
|
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
|||
|
|
- lr_scheduler_type: linear
|
|||
|
|
- lr_scheduler_warmup_steps: 100
|
|||
|
|
- num_epochs: 4
|
|||
|
|
|
|||
|
|
### Training results
|
|||
|
|
|
|||
|
|
| Training Loss | Epoch | Step | Validation Loss |
|
|||
|
|
|:-------------:|:-----:|:-----:|:---------------:|
|
|||
|
|
| 2.3108 | 1.0 | 7500 | 2.1799 |
|
|||
|
|
| 2.265 | 2.0 | 15000 | 2.1632 |
|
|||
|
|
| 2.2507 | 3.0 | 22500 | 2.1567 |
|
|||
|
|
| 2.2519 | 4.0 | 30000 | 2.1534 |
|
|||
|
|
|
|||
|
|
|
|||
|
|
### Framework versions
|
|||
|
|
|
|||
|
|
- Transformers 4.31.0
|
|||
|
|
- Pytorch 2.0.1+cu118
|
|||
|
|
- Datasets 2.13.1
|
|||
|
|
- Tokenizers 0.13.3
|