|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
tags: |
|
- t5 |
|
- qa |
|
- askscience |
|
- lfqa |
|
- information retrieval |
|
datasets: |
|
- vblagoje/lfqa |
|
metrics: |
|
- rouge |
|
widget: |
|
- text: "why aren't there more planets in our solar system?" |
|
example_title: "solar system" |
|
- text: "question: what is a probability distribution? context: I am just learning about statistics." |
|
example_title: "probability distribution" |
|
- text: "question: What are the underlying physical processes by which exercise helps us lose weight? context: I started working out two weeks ago and already feel a lot better, and started to think about it and became deeply confused." |
|
example_title: "pumpen" |
|
- text: "what is a neural network?" |
|
example_title: "deep learning" |
|
- text: "What are the primary mechanisms that computers use to understand human language?" |
|
example_title: "NLP" |
|
|
|
inference: |
|
parameters: |
|
max_length: 96 |
|
no_repeat_ngram_size: 2 |
|
encoder_no_repeat_ngram_size: 4 |
|
repetition_penalty: 3.51 |
|
length_penalty: 0.8 |
|
num_beams: 4 |
|
early_stopping: True |
|
|
|
--- |
|
|
|
<!-- 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. --> |
|
|
|
# checkpoints |
|
|
|
- This model is a fine-tuned version of [google/t5-v1_1-base](https://huggingface.co/google/t5-v1_1-base) on the `vblagoje/lfqa` dataset, with training duration of 2 epochs, for a (_somewhat_) apples-to-apples comparison with [t5-base](https://huggingface.co/pszemraj/t5-base-askscience) on the standard eli5 dataset. |
|
- This checkpoint does seem to be more coherent than t5-base on the original dataset. |
|
- Compared to [bart on lfqa](https://huggingface.co/vblagoje/bart_lfqa), it seems to be able to respond to some questions independently of retrieval. |
|
|
|
## Model description |
|
|
|
More information needed |
|
|
|
## Intended uses & limitations |
|
|
|
- Q&A, information retrieval |
|
- it is probably better to use it with a [retrieval pipeline](https://github.com/deepset-ai/haystack) than alone |
|
|
|
## Training and evaluation data |
|
|
|
More information needed |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 4e-05 |
|
- train_batch_size: 8 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- distributed_type: multi-GPU |
|
- gradient_accumulation_steps: 2 |
|
- total_train_batch_size: 16 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: cosine |
|
- num_epochs: 2 |
|
|
|
### Training results |
|
|
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.16.2 |
|
- Pytorch 1.10.0+cu113 |
|
- Datasets 1.18.3 |
|
- Tokenizers 0.11.0 |
|
|
