ArpitJha/Indian-FinBert-Adapters

Indian FinBERT — LoRA Fine-tuned Adapter

FinBERT (ProsusAI/finbert) fine-tuned on Indian financial news headlines using LoRA adapters.
Optimised for Indian market sentiment — Nifty 50 stocks, NSE/BSE news, RBI announcements, and Indian business headlines. This repo contains the adapter files only .

---

The adapter is ~15 MB — not the full 440 MB model. You load it on top of the base ProsusAI/finbert model which is downloaded automatically from HuggingFace on first use.

---

Labels

ID	Label	Meaning
0	POSITIVE	Bullish sentiment
1	NEGATIVE	Bearish sentiment
2	NEUTRAL	No clear directional signal

---

Setup

1. Clone the repo and unzip the adapter

# Make sure git-xet is installed (https://hf.co/docs/hub/git-xet)
winget install git-xet
git clone https://huggingface.co/ArpitJha/Indian-FinBert-Adapters

Or just download the files and move them in a single folder for ease

2. Create a virtual environment

python -m venv .venv

# Linux / macOS
source .venv/bin/activate

# Windows
.venv\Scripts\activate

3. Install dependencies

pip install torch==2.6.0 transformers==4.47.0 peft==0.13.0 safetensors==0.4.3

Have an NVIDIA GPU? Install the CUDA build of torch for faster inference:

pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install transformers==4.47.0 peft==0.13.0 safetensors==0.4.3

---

Option A — Quick inference with the adapter

Load the adapter on top of the frozen base model. This is the lightest approach — base model weights stay on disk, only the adapter is in memory on top.

from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    pipeline
)
from peft import PeftModel
import torch

BASE_MODEL   = "ProsusAI/finbert"

# Locally after downloading and unzipping the adapter files, we can load them using the path
ADAPTER_PATH = "YOUR_ADAPTER_FOLDER"


tokenizer  = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForSequenceClassification.from_pretrained(BASE_MODEL, num_labels=3)

model = PeftModel.from_pretrained(base_model, ADAPTER_PATH)
model = model.merge_and_unload()
model.eval()

nlp = pipeline(
    "sentiment-analysis",
    model     = model,
    tokenizer = tokenizer,
    device    = 0 if torch.cuda.is_available() else -1
)

label_map = {
    "LABEL_0": "POSITIVE",
    "LABEL_1": "NEGATIVE",
    "LABEL_2": "NEUTRAL"
}

test_headlines = [
"Reliance Industries posts record quarterly profit.",
    "Adani Group stocks crash amid fraud allegations.",
    "RBI keeps interest rates unchanged in policy meeting.",
    "Infosys wins $2 billion AI transformation deal.",
]

print("=" * 60)
print("FINBERT + LORA SENTIMENT ANALYSIS")
print("=" * 60)

for text in test_headlines:
    result    = nlp(text)[0]
    sentiment = label_map.get(result["label"], result["label"])
    print(f"\nHeadline : {text}")
    print(f"Sentiment: {sentiment}")
    print(f"Confidence: {result['score']:.4f}")

---

Requirements

torch>=2.6.0
transformers==4.47.0
peft==0.13.0
safetensors==0.4.3
pandas>=2.2.0          # only needed for batch CSV inference

---

Model details

Property	Value
Base model	ProsusAI/finbert
Fine-tuning method	LoRA (PEFT)
LoRA rank	32
LoRA alpha	64
Target modules	query, value
Training data	Indian financial news headlines
Task	3-class sentiment classification
Labels	Positive / Negative / Neutral
Trainable parameters	~1% of total

---

License

The adapter weights in this repository are released under the MIT License.
The base model (ProsusAI/finbert) is subject to its own Apache 2.0 license.