Training procedure
The following bitsandbytes quantization config was used during training:
- load_in_8bit: True
- load_in_4bit: False
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: fp4
- bnb_4bit_use_double_quant: False
- bnb_4bit_compute_dtype: float32
Model Description
For more information on how it was created, check out the following link: https://github.com/DunnBC22/NLP_Projects/blob/main/OPT%20Models/Grade%20School%20Math%20Instructions%20Fine-Tune%20OPT.ipynb.
Intended uses & limitations
This is intended to show the possibilities. It is mainly limited by the input data.
Training & Evaluation Dataset
Dataset Source: https://huggingface.co/datasets/qwedsacf/grade-school-math-instructions
Hyperparameters Used
| Hyperperameter | Value |
|---|---|
| Model Checkpoint | facebook/opt-2.7b |
| per_device_train_batch_size | 4 |
| gradient_accumulation_steps | 4 |
| fp16 | True |
| warmup_steps | 225 |
| learning_rate | 2e-4 |
| Training Steps | 450 |
Framework versions
| Library | Version |
|---|---|
| Python | 3.10.1 |
| Torch | 2.0.1+cu118 |
| Datasets | 2.14.4 |
| Transformer | 4.31.0 |
| PEFT | 0.4.0 |
Metric
Perplexity = 6.35
License
This model is a fine-tuned version of Meta's OPT-2.7B model.
The original model is released under a custom license (Meta OPT license). The training dataset (grade-school-math-instructions) is derived from GSM8K and may have its own licensing terms.
As a result, this model is released under a non-standard license. Users must review and comply with all upstream licenses.
License Notice
This model is a fine-tuned derivative of a pretrained model. Users must comply with the original model license.
Dataset Notice
This model was fine-tuned on third-party datasets which may have separate licenses or usage restrictions.
- Downloads last month
- 20