Edit model card

bart-base-instructiongen

Instead of generating questions from text, generate instructions for LLMs!

About

Hypothesis: Apply text-to-text models to unlabeled domain-specific text to generate appropriate LLM instructions. Consequently, this may enable domain adaptation of instruction-tuned LLMs, making them more versatile for specific domains.

This model is a fine-tuned version of the facebook/bart-base model, fine-tuned using the pszemraj/fleece2instructions dataset.

It achieves the following results on the evaluation set:

  • Loss: 1.0034
  • Rouge1: 61.7209
  • Rouge2: 45.0116
  • Rougel: 59.8188
  • Rougelsum: 59.8931
  • Gen Len: 14.3179

Intended uses & limitations

This is just a base model/example. There is likely to be even better performance with larger models (click here to see other checkpoints)

Additionally, this was trained on a dataset of only instructions+outputs, with the inputs filtered out. This means that text of 1) cookies and cream 2) chocolate chip 3) mint chip 4) oreo will not get you "Rank the following ice cream flavors: oreo, mint chip, chocolate chip, cookies and cream".

Training and evaluation data

See the linked dataset pszemraj/fleece2instructions - it is a filtered/formatted version of tatsu-lab/alpaca to generate instructions for arbitrary text.

  • Some of the API examples are intentionally weird to demonstrate the generalizability of the model.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 8e-05
  • train_batch_size: 8
  • eval_batch_size: 1
  • seed: 42
  • distributed_type: multi-GPU
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 64
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_ratio: 0.02
  • num_epochs: 2.0

Training results

Training Loss Epoch Step Validation Loss Rouge1 Rouge2 Rougel Rougelsum Gen Len
1.2723 1.0 362 1.0325 61.6206 45.1199 59.6467 59.7534 14.0443
1.0157 2.0 724 1.0034 62.4433 46.0114 60.5355 60.6392 14.1807
Downloads last month
16
Safetensors
Model size
139M params
Tensor type
F32
·

Dataset used to train pszemraj/bart-base-instructiongen

Space using pszemraj/bart-base-instructiongen 1

Evaluation results