checkpoints

• This model is a fine-tuned version of 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 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, it seems to be able to respond to some questions independently of retrieval.

NOTE: the inference API is limited to generating approx. 64 chars for runtime reasons, for longer outputs try using it in python as a transformers pipeline object.

Intended uses & limitations

• Q&A, information retrieval
• it is probably better to use it with a retrieval pipeline than alone

Training and evaluation data

• see linked dataset. the dataset was filtered to only included the askscience subreddit in an attempt to focus on academic/technical queries.

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