|
--- |
|
license: gemma |
|
library_name: peft |
|
tags: |
|
- trl |
|
- sft |
|
- generated_from_trainer |
|
base_model: google/gemma-2b |
|
datasets: |
|
- generator |
|
model-index: |
|
- name: gemma-dolly-agriculture |
|
results: [] |
|
--- |
|
|
|
# gemma-dolly-agriculture |
|
|
|
This model is based on [google/gemma-2b](https://huggingface.co/google/gemma-2b), fine tuned with the dolly-qa dataset and some specific examples of agricultural disease descriptions. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 2.0198 |
|
|
|
## How to Run Inference |
|
|
|
``` |
|
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer |
|
|
|
model_id = "google/gemma-2b" |
|
peft_model_id = "apfurman/gemma-dolly-agriculture" |
|
|
|
# make sure you have access to gemma-2b as well |
|
model = AutoModelForCausalLM.from_pretrained(model_id, token="YOUR_TOKEN_HERE") |
|
model.load_adapter(peft_model_id) |
|
tokenizer = AutoTokenizer.from_pretrained(model_id, token="YOUR_TOKEN_HERE") |
|
|
|
def ask(prompt): |
|
inputs = tokenizer(prompt, return_tensors="pt").input_ids |
|
with torch.inference_mode(): |
|
tokens = model.generate( |
|
inputs, |
|
pad_token_id=128001, |
|
eos_token_id=128001, |
|
max_new_tokens=200, |
|
repetition_penalty=1.5, |
|
) |
|
|
|
return tokenizer.decode(tokens[0], skip_special_tokens=True) |
|
``` |
|
|
|
## Intended uses & limitations |
|
|
|
Created for prompting an AI about agricultural info, but more fine-tuning is needed as current results are not great. |
|
|
|
## Training and evaluation data |
|
|
|
|
|
## Training procedure |
|
|
|
Trained on Intel Data Center GPU Max Series with Intel Developer Cloud running a jupyter notebook. |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 1e-05 |
|
- train_batch_size: 2 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 8 |
|
- total_train_batch_size: 16 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- lr_scheduler_warmup_ratio: 0.05 |
|
- training_steps: 1480 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | |
|
|:-------------:|:-------:|:----:|:---------------:| |
|
| 2.918 | 1.6393 | 100 | 2.5702 | |
|
| 2.4342 | 3.2787 | 200 | 2.2747 | |
|
| 2.2482 | 4.9180 | 300 | 2.1601 | |
|
| 2.1554 | 6.5574 | 400 | 2.0971 | |
|
| 2.1022 | 8.1967 | 500 | 2.0698 | |
|
| 2.0806 | 9.8361 | 600 | 2.0544 | |
|
| 2.0651 | 11.4754 | 700 | 2.0437 | |
|
| 2.0439 | 13.1148 | 800 | 2.0359 | |
|
| 2.0369 | 14.7541 | 900 | 2.0302 | |
|
| 2.034 | 16.3934 | 1000 | 2.0263 | |
|
| 2.0249 | 18.0328 | 1100 | 2.0236 | |
|
| 2.0174 | 19.6721 | 1200 | 2.0218 | |
|
| 2.0154 | 21.3115 | 1300 | 2.0203 | |
|
| 2.0145 | 22.9508 | 1400 | 2.0198 | |
|
|
|
|
|
### Framework versions |
|
|
|
- PEFT 0.10.0 |
|
- Transformers 4.40.1 |
|
- Pytorch 2.1.0.post0+cxx11.abi |
|
- Datasets 2.19.0 |
|
- Tokenizers 0.19.1 |
