pszemraj/gpt2-medium-vaguely-human-dialogue
This model is a fine-tuned version of gpt2-medium on a parsed version of Wizard of Wikipedia. Because the batch size was so large, it learned a general understanding of words that makes sense together but does not specifically respond to anything - sort of like an alien learning to imitate human words to convince others that it is human.
It achieves the following results on the evaluation set:
- Loss: 4.3281
Model description
- a decent example of what happens when your batch size is too large and the global optima does not reflect specific prompts / use cases.
Intended uses & limitations
- there are no intended uses
Training and evaluation data
- a parsed version of the wizard of Wikipedia dataset
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 2
- 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: 0.05
- num_epochs: 10
Training results
Training Loss | Epoch | Step | Validation Loss |
---|---|---|---|
34.991 | 1.0 | 837 | 14.8359 |
12.2881 | 2.0 | 1674 | 9.375 |
8.5071 | 3.0 | 2511 | 7.2148 |
7.6031 | 4.0 | 3348 | 6.1758 |
6.4808 | 5.0 | 4185 | 5.5820 |
5.8562 | 6.0 | 5022 | 5.0977 |
5.6094 | 7.0 | 5859 | 4.8203 |
5.2591 | 8.0 | 6696 | 4.5977 |
5.0031 | 9.0 | 7533 | 4.4219 |
4.8837 | 10.0 | 8370 | 4.3281 |
Framework versions
- Transformers 4.16.1
- Pytorch 1.10.0+cu111
- Tokenizers 0.11.0
- Downloads last month
- 27
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.