little-llama2-ft-qa / README.md
pedromatias97's picture
Update Model Card with inference script
ce94105 verified
---
license: apache-2.0
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
tags:
- generated_from_trainer
model-index:
- name: little-llama2-ft-qa
results: []
library_name: peft
---
<!-- 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. -->
# little-llama2-ft-qa
This model is a fine-tuned version of [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) on an unknown dataset.
It achieves the following results on the evaluation set:
- Loss: 1.5732
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
The following `bitsandbytes` quantization config was used during training:
- quant_method: bitsandbytes
- _load_in_8bit: False
- _load_in_4bit: True
- 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: nf4
- bnb_4bit_use_double_quant: False
- bnb_4bit_compute_dtype: float16
- load_in_4bit: True
- load_in_8bit: False
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.05
- num_epochs: 1
### Training results
| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 1.2789 | 0.2 | 250 | 1.5908 |
| 0.9655 | 0.4 | 500 | 1.5828 |
| 0.9788 | 0.6 | 750 | 1.5764 |
| 1.3064 | 0.8 | 1000 | 1.5739 |
| 1.0251 | 1.0 | 1250 | 1.5732 |
## Inference
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
# prompt template
def generate_inference_prompt(context, question):
return f"""### Instruction: Please answer to the question based on the context information provided. If you don't know the answer, please just say you don't know it, don't try to make an answer from that.\n
### Context:
{context.strip()}\n
### Question:
{question.strip()}
### Answer:
""".strip()
# context to answer
context = """
Great Britain (commonly shortened to Britain) is an island in the North Atlantic Ocean off the north-west coast of continental Europe, consisting of England, Scotland and Wales. With an area of 209,331 km2 (80,823 sq mi), it is the largest of the British Isles, the largest European island and the ninth-largest island in the world. It is dominated by a maritime climate with narrow temperature differences between seasons. The island of Ireland, with an area 40 per cent that of Great Britain, is to the west—these islands, along with over 1,000 smaller surrounding islands and named substantial rocks, form the British Isles archipelago.
"""
# question to ask
question = """
What is the % of area occupied by Ireland in Great Britain?
"""
# loading model
model = AutoModelForCausalLM.from_pretrained(
'pedromatias97/little-llama2-ft-qa'
)
# load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
'pedromatias97/little-llama2-ft-qa'
)
# pipeline
pipe = pipeline(
"text-generation",
model=model,
tokenizer = tokenizer,
torch_dtype=torch.bfloat16,
device_map="auto"
)
# generate prompt
prompt = generate_inference_prompt(context, question)
# generate text
sequences = pipe(
prompt,
do_sample=True,
max_new_tokens=10,
temperature=0.7,
top_k=50,
top_p=0.95,
num_return_sequences=1,
)
# print result
print(sequences[0]['generated_text'])
### output: 40 per cent that of Great Britain
```
### Framework versions
- PEFT 0.5.0
- Transformers 4.38.2
- Pytorch 2.2.1+cu121
- Datasets 2.19.0
- Tokenizers 0.15.2