| # What is SamSum Bot? | |
| This is a model fine-tuned on the [SamSum dataset](https://huggingface.co/datasets/samsum). | |
| However, instead of training the system to summarize conversations, the model is trained to predict a conversation given a summary. | |
| The prompt needs to be in the following form | |
| ```python | |
| A partial summary of the conversation is: | |
| {summary} | |
| With the dialogue being: | |
| {dialogue} | |
| ``` | |
| where *{summary}* is a text as in | |
| ```python | |
| John went out to buy groceries. He meets Jane on the way and they talk about the weather. | |
| ``` | |
| and the *{dialogue}* needs to be structured with speaking lines preceded by the speaking character | |
| ```python | |
| John: Oh hi Jane. | |
| Jane: Nice to see you? | |
| John: The weather looks nice today | |
| Jane: [PREDICTION] | |
| ``` | |
| The system is based on the GPTJ-6B by EleutherAI, [quantized by Hivemind](https://huggingface.co/hivemind/gpt-j-6B-8bit). It has been fine-tuned according to the [LoRa method](https://arxiv.org/abs/2106.09685). | |
| A simple back-end is available in [this repo](https://github.com/fractalego/samsum-bot), where the model is served using Torchserve. | |
| A terminal-like front-end interface is available [here](https://github.com/fractalego/samsumbot_client). | |
| This interface is the one used in my website [http://fractalego.io](http://fractalego.io). | |