|
|
--- |
|
|
license: mit |
|
|
datasets: |
|
|
- virattt/financial-qa-10K |
|
|
language: |
|
|
- en |
|
|
metrics: |
|
|
- accuracy |
|
|
base_model: |
|
|
- EleutherAI/pythia-410m |
|
|
pipeline_tag: text-generation |
|
|
tags: |
|
|
- finance |
|
|
--- |
|
|
--- |
|
|
|
|
|
-- |
|
|
|
|
|
|
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Descriptio |
|
|
|
|
|
- **Developed by:** [Haq Nawaz Malik] |
|
|
- **Model type:** [Lora_adapter] |
|
|
- **Language(s) (NLP):** [Text_gen_for_financial_purposes] |
|
|
- **Finetuned from model :** [EleutherAI/pythia-410m] |
|
|
|
|
|
] |
|
|
### Direct Use |
|
|
|
|
|
``` |
|
|
|
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
from peft import PeftModel |
|
|
import torch |
|
|
|
|
|
# Load tokenizer |
|
|
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-410m") |
|
|
tokenizer.pad_token = tokenizer.eos_token |
|
|
|
|
|
# Load base model |
|
|
base_model = AutoModelForCausalLM.from_pretrained("EleutherAI/pythia-410m").eval().to("cuda" if torch.cuda.is_available() else "cpu") |
|
|
|
|
|
# Load LoRA fine-tuned adapter from Hugging Face Hub |
|
|
lora_model = PeftModel.from_pretrained( |
|
|
base_model, |
|
|
"Omarrran/pythia-financial-lora" |
|
|
).eval().to(base_model.device) |
|
|
|
|
|
# Define prompt |
|
|
prompt = "### Instruction:\n What are Tesla's main risk factors?\n\n### Response:\n" |
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(base_model.device) |
|
|
|
|
|
# Generate from base model |
|
|
with torch.no_grad(): |
|
|
base_output = base_model.generate( |
|
|
**inputs, |
|
|
max_new_tokens=1000, |
|
|
do_sample=True, |
|
|
temperature=0.7, |
|
|
top_p=0.95, |
|
|
repetition_penalty=1.1, |
|
|
eos_token_id=tokenizer.eos_token_id |
|
|
) |
|
|
|
|
|
# Generate from fine-tuned model |
|
|
with torch.no_grad(): |
|
|
lora_output = lora_model.generate( |
|
|
**inputs, |
|
|
max_new_tokens=1000, |
|
|
do_sample=True, |
|
|
temperature=0.7, |
|
|
top_p=0.95, |
|
|
repetition_penalty=1.1, |
|
|
eos_token_id=tokenizer.eos_token_id |
|
|
) |
|
|
|
|
|
# Decode responses |
|
|
base_text = tokenizer.decode(base_output[0], skip_special_tokens=True) |
|
|
lora_text = tokenizer.decode(lora_output[0], skip_special_tokens=True) |
|
|
|
|
|
# Clean output (remove prompt part) |
|
|
base_response = base_text.split("### Response:")[-1].strip() |
|
|
lora_response = lora_text.split("### Response:")[-1].strip() |
|
|
|
|
|
# Display both outputs |
|
|
print("\n" + "="*80) |
|
|
print("📉 BEFORE Fine-Tuning (Base Pythia Model)") |
|
|
print("="*80) |
|
|
print(format_response(base_response)) |
|
|
|
|
|
print("\n" + "="*80) |
|
|
print("📈 AFTER Fine-Tuning (LoRA Adapter from Hugging Face)") |
|
|
print("="*1180) |
|
|
print(format_response(lora_response)) |
|
|
|
|
|
|
|
|
``` |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
 |
|
|
|
|
|
<!-- These are the evaluation metrics being used, ideally with a description of why. --> |
|
|
|
|
|
 |
|
|
|
|
|
|
|
|
 |
