Constant FinSense 0.5 — 140 KB Financial Sentiment Analysis

A 3,109x compressed financial sentiment classifier distilled from FinBERT. Classifies financial text as positive, negative, or neutral in 140 KB.

FinSense vs FinBERT

Model	Accuracy	Size	Latency (CPU)	Compression
FinBERT (ProsusAI)	97.9%	420 MB	~8ms (GPU)	1x
FinSense FP32	84.2%	0.47 MB	0.12ms	894x
FinSense INT8	84.4%	0.14 MB	0.58ms	3,109x

84% accuracy at 140 KB. Runs on microcontrollers. No GPU required.

Quick Start

import onnxruntime as ort
import numpy as np

# Load the INT8 model (140 KB)
session = ort.InferenceSession("finsense_int8.onnx")

# Input: token IDs as int64, shape [batch, 128]
# Use the included vocab_map.json for tokenization
input_name = session.get_inputs()[0].name
dummy = np.zeros((1, 128), dtype=np.int64)
logits = session.run(None, {input_name: dummy})[0]

labels = ["negative", "neutral", "positive"]
pred = labels[np.argmax(logits[0])]
print(pred)

With Tokenization

import json
import onnxruntime as ort
import numpy as np
from transformers import AutoTokenizer

# Load FinBERT tokenizer (shared vocabulary base)
tokenizer = AutoTokenizer.from_pretrained("ProsusAI/finbert")

# Load pruned vocabulary map
with open("vocab_map.json") as f:
    vmap = json.load(f)
    vocab_map = {int(k): v for k, v in vmap["map"].items()}

def tokenize(text, max_length=128):
    tokens = tokenizer.tokenize(text)
    ids = tokenizer.convert_tokens_to_ids(tokens)
    # Remap to pruned vocabulary
    unk = vocab_map.get(tokenizer.unk_token_id, 1)
    cls = vocab_map.get(tokenizer.cls_token_id, unk)
    sep = vocab_map.get(tokenizer.sep_token_id, unk)
    remapped = [cls] + [vocab_map.get(i, unk) for i in ids[:max_length-2]] + [sep]
    remapped += [0] * (max_length - len(remapped))
    return np.array([remapped], dtype=np.int64)

# Run inference
session = ort.InferenceSession("finsense_int8.onnx")
labels = ["negative", "neutral", "positive"]

texts = [
    "Revenue increased 15% year-over-year driven by strong demand",
    "The company reported a significant decline in operating margins",
    "Management maintained its full-year guidance unchanged",
]

for text in texts:
    ids = tokenize(text)
    logits = session.run(None, {"input_ids": ids})[0]
    probs = np.exp(logits) / np.exp(logits).sum(axis=1, keepdims=True)
    pred = labels[np.argmax(probs[0])]
    conf = float(probs[0].max())
    print(f"{pred} ({conf:.0%}): {text}")

Architecture

NanoCNN — the same architecture that powers Constant Edge 0.5, adapted for 3-class financial sentiment:

Embedding: 32-dimensional, pruned vocabulary (2,729 financial tokens)
Convolutions: 4 parallel Conv1d banks (filter sizes 2, 3, 4, 5), 64 filters each
Bottleneck: Linear compression (256 → 16)
Classifier: 16 → 48 → 3 (negative / neutral / positive)
Parameters: 121,971
Quantization: INT8 (post-training, ONNX Runtime)

Distillation Pipeline

ProsusAI/FinBERT (97.9%, 420 MB)
    → Knowledge Distillation (T=10.77, α=0.30)
        → NanoCNN Student (84.4%, 0.14 MB)

Teacher: ProsusAI/finbert, evaluated at 97.94% on Financial PhraseBank (AllAgree split)
Distillation: Optuna-optimized (20 trials) temperature and mixing weight
Evaluation: 5-seed mean with standard deviations

Detailed Results

5-Seed Evaluation (FP32)

Seed	Test Accuracy	F1
42	83.53%	0.828
123	85.59%	0.850
456	84.71%	0.843
789	84.41%	0.841
1024	82.65%	0.819
Mean	84.18% ± 1.01%	0.836

INT8 Classification Report

Class	Precision	Recall	F1	Support
Negative	0.800	0.696	0.744	46
Neutral	0.886	0.933	0.909	209
Positive	0.750	0.706	0.727	85
Weighted Avg	0.841	0.844	0.841	340

Inference Benchmarks

Model	Mean Latency	P95 Latency	Throughput
FP32 (0.47 MB)	0.12ms	0.14ms	~8,300/sec
INT8 (0.14 MB)	0.58ms	0.71ms	~1,700/sec

Use Cases

Algorithmic trading — sub-millisecond sentiment scoring on financial news feeds, no GPU required
Bloomberg/Refinitiv terminals — on-device sentiment without API calls or data leaving the terminal
Mobile trading apps — real-time sentiment on earnings releases, SEC filings, analyst reports
Compliance monitoring — on-premise sentiment analysis for MiFID II / SEC audit requirements
IoT financial displays — sentiment-aware tickers on embedded hardware

Training Details

Dataset: Financial PhraseBank (AllAgree), 2,264 sentences from financial news
Split: 70/15/15 stratified (1,584 train / 340 val / 340 test)
Teacher: ProsusAI/finbert (BERT-base fine-tuned on financial sentiment)
Hardware: NVIDIA RTX 4090 Laptop GPU (16 GB)
Optuna: 20 trials, median pruner, best α=0.30, T=10.77
Training: 30 epochs per seed, AdamW, cosine schedule with linear warmup

Model Files

File	Size	Description
`finsense_int8.onnx`	140 KB	INT8 quantized (recommended)
`finsense_fp32.onnx`	470 KB	FP32 full precision
`vocab_map.json`	38 KB	Pruned vocabulary mapping
`results.json`	3 KB	Full benchmark results

Citation

@misc{constantone2026finsense,
  title={FinSense: 3109x Compressed Financial Sentiment Analysis for Edge Deployment},
  author={ConstantOne AI},
  year={2026},
  url={https://huggingface.co/ConstantQJ/constant-finsense-0.5}
}

ConstantQJ
/

constant-finsense-0.5