yakul259/fint5-financeqa-customised

---

🧾 Model Card: fint5-financeqa

🔎 Overview

Model Name: fint5-financeqa Base Model: t5-small Domain: Financial Question Answering Fine-tuned by: [Your Name] Training Environment: Google Colab License: MIT (inherits from t5-small)

---

📚 Description

fint5-financeqa is a fine-tuned version of the t5-small model, tailored for financial question-answering tasks using the sweatSmile/FinanceQA dataset.

The model is capable of answering domain-specific questions when given both the question and relevant context. This makes it suitable for chatbot assistants, enterprise finance automation, or as a component in investment-related QA systems.

---

📦 Intended Use

You can use this model to:

• Build finance-domain QA bots
• Help users query financial reports
• Perform automated FAQ answering for fintech tools

Not recommended for:

• General-purpose QA tasks (use base T5 instead)
• Time-sensitive financial forecasting or advice
• Legal or regulatory decision-making

---

🧠 Training Details

Detail	Value
Base Model	t5-small
Dataset	sweatSmile/FinanceQA
Dataset Size	3705 total entries (80/20 split)
Preprocessing	QUERY + context: CONTEXT
Target Text	ANSWER string
Max Input Length	512 tokens
Max Target Length	128 tokens
Epochs	3
Batch Size	4
Optimizer & LR	AdamW @ 2e-5
Mixed Precision	Enabled (fp16)
Evaluation Strategy	Per Epoch
Frameworks	🤗 Transformers, Datasets, Accelerate

---

📈 Metrics

Metric	Value
Final Eval Loss	0.415
BLEU / ROUGE	[Not explicitly printed]
EM / Accuracy	[Can be added]

Note: Evaluation beyond loss was mentioned but not shown. Future versions can include BLEU, ROUGE, or exact match scores using evaluate or sacrebleu.

---

📉 Training Curves

Loss chart from the training loop:

Replace with actual matplotlib plot or embed image if hosted.

---

🧪 Example Usage

python from transformers import T5Tokenizer, T5ForConditionalGeneration

model = T5ForConditionalGeneration.from_pretrained("./fint5-financeqa") tokenizer = T5Tokenizer.from_pretrained("./fint5-financeqa")

question = "What is the net income of the company?" context = "The company reported a net income of $1.5 million for the last quarter."

input_text = f"{question} context: {context}" input_ids = tokenizer(input_text, return_tensors="pt").input_ids

output = model.generate(input_ids) answer = tokenizer.decode(output[0], skip_special_tokens=True)

print("Answer:", answer)

---

🌱 Carbon Footprint (Optional)

Tracked using codecarbon:

Metric	Value
Emissions (kg CO₂eq)	Reported via emissions.csv
Duration	~[training time]
Hardware	Google Colab GPU

---

🧭 Future Improvements

• Include BLEU / ROUGE / EM scores for comprehensive evaluation
• Apply to other finance datasets (FinQA, Doc2Dial, etc.)
• Use larger base model (e.g., t5-base) for better performance
• Integrate LoRA/PEFT to optimize memory and speed

---

🏁 Model Access

📁 Directory: ./fint5-financeqa ✅ Includes:

• pytorch_model.bin
• config.json
• tokenizer_config.json
• spiece.model

You can upload this model to Hugging Face using transformers-cli or huggingface_hub.

---