ITYES Financial Sentiment

Experimental portfolio demonstrator · Llama 3.1 8B final candidate

Overview

A compact, locally deployable 8B model for financial sentiment classification. It classifies short English financial statements as Positive, Neutral or Negative and demonstrates controlled evaluation, local deployment and traceable AI engineering.

Technical profile

Model family	Meta Llama 3.1 8B Instruct
Parameters	8 billion
Dataset	Financial PhraseBank · AllAgree
Final evaluation	Separate locked holdout · accessed once
Local deployment	GGUF · Q4_K_M · approximately 4.6 GiB
Output contract	Exactly one sentiment label · 100% valid format

Final evaluation

Metric	Result
Accuracy	89.6%
Macro-F1	0.878
Balanced accuracy	0.865
Negative / Neutral / Positive recall	95.6% / 98.1% / 65.9%

The candidate was selected using validation and challenge instruments separate from the final holdout. No final-test baseline comparison is claimed.

Artifact integrity

SHA-256 cryptographic fingerprints identify the private evaluated artifacts. They are not digital signatures or third-party certifications.

Final model adapter: d004b0a3bb17fff4f2c25a91111a1909936c76f17ad2607dd4267d7b92108ecc
Quantized GGUF artifact: fd9b375f716d37d2f1d12fedd43686fc0323f5fa87e4b51c155569e23a9a8639
Final evaluation output: 2083dabf0591bc401802a3747516279ca605d8e989832b49f59dddf1fe57939d

Known limitations

Evaluation covers short English financial statements, not complete articles, filings, tables, earnings calls or production traffic.
Positive recall is lower than negative and neutral recall; a reduction in losses may be classified as neutral.
Mixed or contradictory signals may reasonably receive the neutral label.
This is not a production financial system and must not be used for investment, trading, credit, compliance or other financial decisions.

Dataset and licensing

Financial PhraseBank · AllAgree. ITYES does not redistribute the dataset or examples. Its use in this limited demonstrator is presented with proper attribution and written permission from the dataset author. Malo et al. (2014), https://doi.org/10.1002/asi.23062. Financial PhraseBank is associated with Aalto University and licensed CC BY-NC-SA 3.0.