pravz/slm_v1
pravz/slm_v1 is a 1.5 billion parameter model based on Qwen2.5-1.5B-Instruct, developed by pravz. It is specifically fine-tuned to read Indian corporate filings (NSE/BSE) and extract structured JSON outputs, including event type, materiality, financial amounts, and a prediction of stock movement. Uniquely, its training targets are derived from measured market reactions rather than LLM-assigned sentiment, making it specialized for financial document analysis and market impact prediction.
Loading preview...
Overview
pravz/slm_v1 is a specialized 1.5 billion parameter language model, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. Its core function is to process Indian corporate filings (NSE/BSE) and generate structured JSON output. This output includes critical financial event details such as event_type, materiality, surprise, extracted amount_inr_cr, and predicted direction, shape, price_path, and volume_path of stock movement.
Key Differentiator
What sets this model apart is its supervision by measured market reaction rather than human or LLM-assigned sentiment labels. Unlike models like FinBERT or FinGPT, pravz/slm_v1's targets are derived directly from actual price movements in the 15 minutes post-filing, market-adjusted against NIFTY 50.
Performance & Limitations
While it outperforms DeepSeek on metrics like balanced accuracy and mover ROC-AUC for predicting stock direction, both models struggle against a constant 'FLAT' prediction baseline. The model's primary strength lies in its feature extraction capabilities, where its embeddings, when combined with market features, perform comparably to TF-IDF + SVD. However, the training objective (next-token prediction over a JSON schema) optimized for format compliance, not necessarily for encoding market-moving signals within its hidden states.
Training Details
The model was full fine-tuned on 146,500 NSE filings up to April 2026, using a single epoch. The training process cost approximately $33 on an A10G GPU. The prompt format is crucial for optimal performance, requiring a specific system and user layout as demonstrated in the usage examples.
Intended Use
This model is intended for research and education in financial document analysis. It is explicitly not investment advice and is limited to Indian filings from its training period. Its generative output (event classification, amount extraction) remains untested for accuracy, and its ability to predict market direction is limited, as indicated by benchmarks.