85 lines
2.8 KiB
Markdown
85 lines
2.8 KiB
Markdown
|
---
|
||
|
|
license: mit
|
||
|
|
datasets:
|
||
|
|
- databricks/databricks-dolly-15k
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
# GPT-2-dolly
|
||
|
|
|
||
|
|
**GPT-2-dolly** is an instruction fine-tuned model based on the GPT-2 transformer architecture.
|
||
|
|
|
||
|
|
|
||
|
|
### Benchmark Metrics
|
||
|
|
|
||
|
|
| Metric | GPT-2-dolly | GPT-2 (base) |
|
||
|
|
|-----------------------|-------|-------|
|
||
|
|
| Avg. | **30.91** | 29.99 |
|
||
|
|
| ARC (25-shot) | **22.70** | 21.84 |
|
||
|
|
| HellaSwag (10-shot) | 30.15 | **31.6** |
|
||
|
|
| MMLU (5-shot) | 25.81 | **25.86** |
|
||
|
|
| TruthfulQA (0-shot) | **44.97** | 40.67 |
|
||
|
|
|
||
|
|
|
||
|
|
We use state-of-the-art [Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) to run the benchmark tests above, using the same version as the HuggingFace LLM Leaderboard. Please see below for detailed instructions on reproducing benchmark results.
|
||
|
|
|
||
|
|
### Model Details
|
||
|
|
|
||
|
|
* **Trained by**: Luiz G A Alves
|
||
|
|
* **Model type:** **GPT-2-dolly** is an auto-regressive language model based on the GPT-2 transformer architecture.
|
||
|
|
* **Language(s)**: English
|
||
|
|
|
||
|
|
### How to use:
|
||
|
|
|
||
|
|
```python
|
||
|
|
# Use a pipeline as a high-level helper
|
||
|
|
>>> from transformers import pipeline
|
||
|
|
>>> pipe = pipeline("text-generation", model="lgaalves/gpt2-dolly")
|
||
|
|
>>> question = "What is a large language model?"
|
||
|
|
>>> answer = pipe(question)
|
||
|
|
>>> print(answer[0]['generated_text'])
|
||
|
|
```
|
||
|
|
|
||
|
|
or, you can load the model direclty using:
|
||
|
|
|
||
|
|
```python
|
||
|
|
# Load model directly
|
||
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
|
|
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained("lgaalves/gpt2-dolly")
|
||
|
|
model = AutoModelForCausalLM.from_pretrained("lgaalves/gpt2-dolly")
|
||
|
|
```
|
||
|
|
|
||
|
|
### Training Dataset
|
||
|
|
|
||
|
|
`lgaalves/gpt2-dolly` trained using the Databricks Dolly dataset [`databricks/databricks-dolly-15k`](https://huggingface.co/datasets/databricks/databricks-dolly-15k).
|
||
|
|
|
||
|
|
|
||
|
|
### Training Procedure
|
||
|
|
|
||
|
|
`lgaalves/gpt2-dolly` was instruction fine-tuned using LoRA on 1 T4 GPU on Google Colab. It took about 1.5 hours to train it.
|
||
|
|
|
||
|
|
|
||
|
|
# Intended uses, limitations & biases
|
||
|
|
|
||
|
|
You can use the raw model for text generation or fine-tune it to a downstream task. The model was not extensively tested and may produce false information. It contains a lot of unfiltered content from the internet, which is far from neutral.
|
||
|
|
|
||
|
|
|
||
|
|
# [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_lgaalves__gpt2-dolly)
|
||
|
|
|
||
|
|
| Metric | Value |
|
||
|
|
|-----------------------|---------------------------|
|
||
|
|
| Avg. | 25.53 |
|
||
|
|
| ARC (25-shot) | 22.7 |
|
||
|
|
| HellaSwag (10-shot) | 30.15 |
|
||
|
|
| MMLU (5-shot) | 25.81 |
|
||
|
|
| TruthfulQA (0-shot) | 44.97 |
|
||
|
|
| Winogrande (5-shot) | 51.46 |
|
||
|
|
| GSM8K (5-shot) | 0.15 |
|
||
|
|
| DROP (3-shot) | 3.45 |
|