This was mostly a test to see what the loss/eval looked like when training on top of Harmonia, and in that sense it was a sterling success, without the "jitter" I experienced training on top of Nethena 20b. Quick testing shows a bit of derpiness, but a nice conversational flow. Overall, this will be helpful in developing additional 20b merges.
NOTES
This model is a fine-tuned version of athirdpath/Harmonia-20B on the HF No Robots dataset. It achieves the following results on the evaluation set:
- Loss: 1.4881
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 3.5e-05
- train_batch_size: 3
- eval_batch_size: 3
- seed: 42
- gradient_accumulation_steps: 3
- total_train_batch_size: 9
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 10
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 1.5598 | 0.55 | 50 | 1.5816 |
| 1.5384 | 1.08 | 100 | 1.5146 |
| 1.5362 | 1.64 | 150 | 1.4972 |
| 1.4234 | 2.17 | 200 | 1.4902 |
| 1.4678 | 2.72 | 250 | 1.4881 |
Framework versions
- Transformers 4.35.2
- Pytorch 2.0.1+cu118
- Datasets 2.15.0
- Tokenizers 0.15.0