Sashank1006

finance-sentiment-mistral-lora

README

License: apache-2.0

📊 Performance

Table with columns: Metric, Base Model, Our Model, Improvement
Metric	Base Model	Our Model	Improvement
Win Rate (our dataset)	44	56	+27% relative
Win Rate (market analysis category)	40	60	+50% relative
Dataset quality grade	E (2.0)	B (8.1)	+305% relative

Training curves showed clean, consistent loss reduction with no overfitting across 4 epochs.

🧠 Model Details

Table with columns: Property, Value
Property	Value
Base model	mistralai/Mistral-7B-Instruct-v0.3
Training method	Supervised Fine-Tuning (SFT) + LoRA
LoRA rank	64
LoRA alpha	128
Target layers	q_proj, k_proj, v_proj, o_proj
Epochs	4
Optimizer	Cosine LR scheduler
Warmup ratio	0.05
Gradient clipping

📁 Dataset

The training dataset is a curated merge of 6 sources totalling ~120,000 raw rows, cleaned and deduplicated down to 20,000 high-quality rows via Adaption's Adaptive Data pipeline.

Sources

Table with columns: Source, Type, Rows (approx)
Source	Type	Rows (approx)
financial_phrasebank (sentences_allagree)	Human-labeled news sentences	~2,200
zeroshot/twitter-financial-news-sentiment	Human-labeled financial tweets	~9,900
flwrlabs/fingpt-sentiment-train	Financial NLP training data	~76,800
TimKoornstra/financial-tweets-sentiment	Human-labeled financial tweets	~38,000
nickmuchi/financial-classification	Financial text classification	~2,000
Hand-labeled originals (Indian market)

What makes this dataset original

Indian market focus — original hand-labeled examples covering NSE, BSE, Sensex, Nifty, RBI decisions, Indian fintech (Paytm, Zomato, PhonePe), and Indian conglomerates (Reliance, Tata, Adani, HDFC)
Multi-source deduplication — priority-aware deduplication ensures highest-quality copy is retained when the same text appears across sources
Multilingual context — includes financial terminology specific to the Indian subcontinent not present in standard Western finance NLP datasets
Live news augmentation — recent business headlines via NewsAPI add temporal diversity beyond static datasets

Label distribution (after cleaning)

text

positive :  ~35%negative :  ~33%neutral  :  ~32%

Balanced across all three classes to prevent label bias.

Data quality improvement via Adaptive Data

Adaption's Adaptive Data pipeline was applied before training:

Before: Grade E, quality score 2.0, percentile 0.1
After: Grade B, quality score 8.1, percentile 17.8
Relative improvement: 305%

🚀 How to Use

python

from transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModel base_model = "mistralai/Mistral-7B-Instruct-v0.3"lora_model  = "Sashank1006/finance-sentiment-mistral-lora" tokenizer = AutoTokenizer.from_pretrained(base_model)model     = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")model     = PeftModel.from_pretrained(model, lora_model)model.eval() def predict_sentiment(text: str) -> str:    prompt = f"Classify the sentiment of this financial text as positive, negative, or neutral:\n\n{text}\n\nSentiment:"    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)    output = model.generate(**inputs, max_new_tokens=5, do_sample=False)    result = tokenizer.decode(output[0], skip_special_tokens=True)    return result.split("Sentiment:")[-1].strip().lower() # Exampletext = "Reliance Industries reported a 23% jump in quarterly profit."print(predict_sentiment(text))  # → "positive"

🌍 Real-World Applications

Retail investor tools — classify financial news before displaying to users
Trading signal generation — convert news sentiment into bullish/bearish signals
Portfolio risk monitoring — flag negative sentiment around held stocks
Indian fintech apps — specifically tuned for Indian market terminology and companies
News aggregators — auto-tag financial articles by sentiment

⚠️ Limitations

Trained primarily on English-language financial text; performance on Hindi/Tamil/regional language finance text will be lower
Rule-labelled NewsAPI headlines (~500 rows) may contain some label noise
Model may underperform on highly technical financial derivative or options-specific language
Sentiment is classified at the sentence/headline level — document-level sentiment aggregation requires additional logic
The model reflects sentiment patterns in training data up to mid-2026; sentiment around newer entities may be less accurate

📋 Training Pipeline

text

Raw data (6 sources, ~120K rows)        ↓Merge + deduplicate (prepare_dataset.py)        ↓Upload to Adaption Adaptive Data        ↓Adaptive Data optimization (Grade E → B, 305% quality improvement)        ↓AutoScientist fine-tuning (Mistral-7B-Instruct, LoRA, 4 epochs)        ↓Evaluation (Win rate: 44 → 56 on dataset, 40 → 60 on market analysis)        ↓Released on HuggingFace + Kaggle

👥 Team

Team Caribou — HackIndia Adaption AutoScientist Challenge, Finance Track

Built using:

Adaption AutoScientist — automated model training
Adaption Adaptive Data — dataset quality optimization
HuggingFace Datasets — source data
NewsAPI — live headline augmentation

📄 Citation

bibtex

@misc{caribou2026financeSentiment,  title  = {Finance Sentiment Classifier — Indian Market Focus},  author = {Team Caribou},  year   = {2026},  url    = {https://huggingface.co/YOUR_HF_USERNAME/finance-sentiment-mistral-lora},  note   = {Built for HackIndia Adaption AutoScientist Challenge}}

🔗 Links

📦 Dataset: Kaggle — finance-sentiment-india
🤗 Model: HuggingFace — finance-sentiment-mistral-lora
🎮 Demo: HuggingFace Spaces — Gradio App

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Sashank1006

Model Tree

Base

mistralai/Mistral-7B-Instruct-v0.3

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

📊 Performance

Table with columns: Metric, Base Model, Our Model, Improvement
Metric	Base Model	Our Model	Improvement
Win Rate (our dataset)	44	56	+27% relative
Win Rate (market analysis category)	40	60	+50% relative
Dataset quality grade	E (2.0)	B (8.1)	+305% relative

Training curves showed clean, consistent loss reduction with no overfitting across 4 epochs.

🧠 Model Details

Table with columns: Property, Value
Property	Value
Base model	mistralai/Mistral-7B-Instruct-v0.3
Training method	Supervised Fine-Tuning (SFT) + LoRA
LoRA rank	64
LoRA alpha	128
Target layers	q_proj, k_proj, v_proj, o_proj
Epochs	4
Optimizer	Cosine LR scheduler
Warmup ratio	0.05
Gradient clipping

📁 Dataset

The training dataset is a curated merge of 6 sources totalling ~120,000 raw rows, cleaned and deduplicated down to 20,000 high-quality rows via Adaption's Adaptive Data pipeline.

Sources

Table with columns: Source, Type, Rows (approx)
Source	Type	Rows (approx)
financial_phrasebank (sentences_allagree)	Human-labeled news sentences	~2,200
zeroshot/twitter-financial-news-sentiment	Human-labeled financial tweets	~9,900
flwrlabs/fingpt-sentiment-train	Financial NLP training data	~76,800
TimKoornstra/financial-tweets-sentiment	Human-labeled financial tweets	~38,000
nickmuchi/financial-classification	Financial text classification	~2,000
Hand-labeled originals (Indian market)

What makes this dataset original

Indian market focus — original hand-labeled examples covering NSE, BSE, Sensex, Nifty, RBI decisions, Indian fintech (Paytm, Zomato, PhonePe), and Indian conglomerates (Reliance, Tata, Adani, HDFC)
Multi-source deduplication — priority-aware deduplication ensures highest-quality copy is retained when the same text appears across sources
Multilingual context — includes financial terminology specific to the Indian subcontinent not present in standard Western finance NLP datasets
Live news augmentation — recent business headlines via NewsAPI add temporal diversity beyond static datasets

Label distribution (after cleaning)

text

positive :  ~35%negative :  ~33%neutral  :  ~32%

Balanced across all three classes to prevent label bias.

Data quality improvement via Adaptive Data

Adaption's Adaptive Data pipeline was applied before training:

Before: Grade E, quality score 2.0, percentile 0.1
After: Grade B, quality score 8.1, percentile 17.8
Relative improvement: 305%

🚀 How to Use

python

from transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModel base_model = "mistralai/Mistral-7B-Instruct-v0.3"lora_model  = "Sashank1006/finance-sentiment-mistral-lora" tokenizer = AutoTokenizer.from_pretrained(base_model)model     = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")model     = PeftModel.from_pretrained(model, lora_model)model.eval() def predict_sentiment(text: str) -> str:    prompt = f"Classify the sentiment of this financial text as positive, negative, or neutral:\n\n{text}\n\nSentiment:"    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)    output = model.generate(**inputs, max_new_tokens=5, do_sample=False)    result = tokenizer.decode(output[0], skip_special_tokens=True)    return result.split("Sentiment:")[-1].strip().lower() # Exampletext = "Reliance Industries reported a 23% jump in quarterly profit."print(predict_sentiment(text))  # → "positive"

🌍 Real-World Applications

Retail investor tools — classify financial news before displaying to users
Trading signal generation — convert news sentiment into bullish/bearish signals
Portfolio risk monitoring — flag negative sentiment around held stocks
Indian fintech apps — specifically tuned for Indian market terminology and companies
News aggregators — auto-tag financial articles by sentiment

⚠️ Limitations

Trained primarily on English-language financial text; performance on Hindi/Tamil/regional language finance text will be lower
Rule-labelled NewsAPI headlines (~500 rows) may contain some label noise
Model may underperform on highly technical financial derivative or options-specific language
Sentiment is classified at the sentence/headline level — document-level sentiment aggregation requires additional logic
The model reflects sentiment patterns in training data up to mid-2026; sentiment around newer entities may be less accurate

📋 Training Pipeline

text

Raw data (6 sources, ~120K rows)        ↓Merge + deduplicate (prepare_dataset.py)        ↓Upload to Adaption Adaptive Data        ↓Adaptive Data optimization (Grade E → B, 305% quality improvement)        ↓AutoScientist fine-tuning (Mistral-7B-Instruct, LoRA, 4 epochs)        ↓Evaluation (Win rate: 44 → 56 on dataset, 40 → 60 on market analysis)        ↓Released on HuggingFace + Kaggle

👥 Team

Team Caribou — HackIndia Adaption AutoScientist Challenge, Finance Track

Built using:

Adaption AutoScientist — automated model training
Adaption Adaptive Data — dataset quality optimization
HuggingFace Datasets — source data
NewsAPI — live headline augmentation

📄 Citation

bibtex

@misc{caribou2026financeSentiment,  title  = {Finance Sentiment Classifier — Indian Market Focus},  author = {Team Caribou},  year   = {2026},  url    = {https://huggingface.co/YOUR_HF_USERNAME/finance-sentiment-mistral-lora},  note   = {Built for HackIndia Adaption AutoScientist Challenge}}

🔗 Links

📦 Dataset: Kaggle — finance-sentiment-india
🤗 Model: HuggingFace — finance-sentiment-mistral-lora
🎮 Demo: HuggingFace Spaces — Gradio App