rroshann/sec-sentiment-sft-deepseek-14b
The rroshann/sec-sentiment-sft-deepseek-14b is a 14.8 billion parameter DeepSeek-R1-Distill-Qwen-14B model, fine-tuned using QLoRA for 5-class sentiment classification of thematic factors from U.S. industrials SEC filings (10-K, 10-Q). Developed as part of an AllianceBernstein × Vanderbilt DSI capstone project, it outputs ordinal sentiment labels, natural-language rationales, and confidence scores. This model is specifically designed for financial-materiality sentiment analysis within SEC-filing prose, not for general-purpose assistance.
Loading preview...
Model Overview
rroshann/sec-sentiment-sft-deepseek-14b is a specialized 14.8 billion parameter model, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-14B using QLoRA. Its primary function is 5-class sentiment classification of thematic factors extracted from U.S. industrials SEC filings (10-K, 10-Q). The model outputs one of five ordinal labels (very_negative, negative, neutral, positive, very_positive), along with a natural-language rationale and a confidence score.
Key Capabilities
- Financial Sentiment Classification: Specifically trained to classify the financial materiality sentiment of individual factor summaries from SEC filings.
- Specialized Training Data: Fine-tuned on 5,217 samples derived from 67,741 thematic factors extracted from 2,441 SEC filings (80 U.S. industrials tickers, 2015-2025), with a two-stage weak-to-strong labeling process involving Claude Opus.
- Validation Accuracy: Achieves 73.3% accuracy on a 1,045-sample held-out validation set (Opus-labeled).
- Portfolio-Level Performance: While per-sample accuracy on realized-return-quintile gold labels is modest (Macro F1 of 0.174), its value is demonstrated in cohort-level aggregation, significantly improving L/S cohort spread in backtests (from 2.78% to 4.88% at 21-day horizon).
Intended Use Cases
This model is in scope for financial-materiality sentiment classification of individual factor summaries from 10-K/10-Q filings. It is not a general-purpose assistant and should not be used for open-ended chat, stock-price prediction, or sentiment analysis outside the U.S. industrials sector or SEC-filing prose. Its optimal use involves cohort-level aggregation and portfolio-level validation as detailed in its technical report.