TA Model 1B / TA Model V2
A released language-model architecture adapted to market data, followed by research into simulation and risk evaluation.
Model trainingPublic repositories: Feb 2026 (1B) · Jun 2026 (V2)
Overview
A transformer trained on structured market data, with a follow-on system for testing decisions before deployment.
Technical work
- Built and released a 1.044B-parameter LLaMA-family decoder (20 layers/20 heads, hidden size 2,000, context 1,860) with a custom 1,024-token vocabulary for structured OHLCVT, price-change fragments, trade/volume buckets, and technical-indicator windows.
- Implemented timestamp normalization, missing-interval repair, contiguous grouping, and 160-step fine-tuning windows with stride 10; published source, model, and dataset. V2 extends the work into an asset-agnostic probabilistic decision/risk platform with lineage, simulation, walk-forward, paper-trading, and promotion gates rather than unsupported prediction claims.