Experimental AI case study
Financial sentiment, evaluated with discipline.
A specialized 8B language model with a quantized local deployment artifact, for classifying short English financial statements as positive, neutral or negative.
338locked holdout examples
01
Why it matters
Financial organizations process thousands of short textual signals—from news headlines and analyst comments to internal reports and customer communications. Reliable sentiment classification can be one building block for systems that help organize, prioritize and analyze this information.
02
Technical highlights
- Task
- Financial sentiment classification
- Labels
- Positive / Neutral / Negative
- Dataset
- Financial PhraseBank · AllAgree
- Evaluation
- Separate locked final holdout
- Metrics
- Accuracy · Macro-F1 · Balanced accuracy
- Output validation
- 100% valid format
Final evaluation
A strong result, presented with its limits.
The selected candidate was evaluated once on a locked holdout of 338 examples. The final evaluation was accessed once; no final-test baseline comparison is claimed.
The candidate was selected on a separate validation split. No baseline was run on the final holdout, and no final-test uplift is claimed.
03
What this demonstrates
Domain adaptation
Adapting a specialized 8B language model for financial language classification.
Evaluation engineering
Building reproducible pipelines with controlled validation and final-test boundaries.
Dataset governance
Keeping source, split, permission and publication decisions traceable.
Responsible communication
Publishing results with uncertainty, limitations and bounded claims.
04
Beyond model adaptation
The project also demonstrates the engineering system around the model: dataset governance, controlled evaluation, experiment reproducibility, publication discipline and responsible AI communication.
- 01Govern
- 02Separate
- 03Validate
- 04Test once
- 05Report
Intended use
This technical portfolio demonstrator showcases a reproducible AI evaluation workflow for specialized financial language.
It is not a production financial system and should not be used for investment, trading, credit or other financial decisions. Results apply only to this dataset, split and protocol.
05
Dataset and attribution
This experiment uses Financial PhraseBank, AllAgree configuration. ITYES does not redistribute the dataset. Its use in this limited demonstrator is presented with proper attribution and written permission from the dataset author.
Financial PhraseBank is associated with Aalto University and licensed under CC BY-NC-SA 3.0. No dataset examples, model weights or adapter files are published here.
Built with Llama. Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
Malo et al. (2014) · Good debt or bad debt ↗About ITYES