|
--- |
|
license: cc-by-nc-4.0 |
|
datasets: |
|
- Dahoas/instruct-synthetic-prompt-responses |
|
language: |
|
- en |
|
pipeline_tag: text-generation |
|
--- |
|
Question answering model finetuned from [GPT4All-J v1.3](https://huggingface.co/nomic-ai/gpt4all-j) with [Direct Preference Optimization](https://arxiv.org/abs/2305.18290). \ |
|
Dataset: [Dahoas/instruct-synthetic-prompt-responses](https://huggingface.co/datasets/Dahoas/instruct-synthetic-prompt-responses). |
|
|
|
The model was finetuned with the following promt: \ |
|
``"Answer the following question in context:\n\nQuestion: " + samples["prompt"] + " Answer: "`` \ |
|
It should be benefical to use the same or a similar prompt for inference. |
|
|
|
An increase in performance compared to [GPT4All-J v1.3](https://huggingface.co/nomic-ai/gpt4all-j) was observed when using two-shot Chain-of-Thought prompting. |
|
|
|
| HellaSwag | WinoGrande | BooLQ | ARC-c | |
|
|:------:|:------:|:------:|:------:| |
|
| 62.37% | 63.3% | 65.2% | 32.76% | |
