Downloads · 30 days
16
34% of all-time downloads
codeparrot-ds
wb-droid/codeparrot-ds
codeparrot-ds is a text generation model from wb-droid. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as mit.
should probably proofread and complete it, then remove this comment. --
All-time downloads
47
Public
Parameters
126M
1 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors503 MB · 99%
At a glance
- Task
- Text Generation
- Library
- transformers
- License
- mit
- Model type
- gpt2
- Access
- Public
- Created
- Sep 29, 2024
- Updated
- Oct 6, 2024
- SHA
- 76add2d9
Try a prompt
Base models
Model card
From the Hugging Face model README
<!-- 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. -->
codeparrot-ds
This model is a fine-tuned version of gpt2 on an unknown dataset. It achieves the following results on the evaluation set:
- Loss: 1.5773
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: 0.0005
- train_batch_size: 8
- eval_batch_size: 18
- seed: 42
- 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_steps: 1000
- num_epochs: 1
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 3.1475 | 0.2329 | 5000 | 2.2790 |
| 2.0924 | 0.4658 | 10000 | 1.9024 |
| 1.7823 | 0.6987 | 15000 | 1.6709 |
| 1.6037 | 0.9316 | 20000 | 1.5773 |
Framework versions
- Transformers 4.45.1
- Pytorch 1.13.1+cu117
- Datasets 3.0.1
- Tokenizers 0.20.0
- Task
- Text Generation
- Library
- transformers
- Type
- gpt2
- License
- mit
- Created
- Sep 29, 2024
- Updated
- Oct 6, 2024