RESEARCH / MACHINE LEARNING
DREAM-001
An exploratory model-state experiment with pre-token features, grouped evaluation and offline reports.
- Current state
- Exploratory prototype
- Discipline
- Research · Machine learning
- Built with
- Python / PyTorch / Hugging Face Transformers / scikit-learn / HTML/JavaScript
The idea
DREAM-001 investigates whether changes in a language model's internal state can precede the first annotated false claim. Its contribution is an explicit extraction and evaluation pipeline that makes timing, labels, folds and uncertainty inspectable.
How it works
Python extracts seven pre-token distribution and hidden-state features from annotated continuations. Grouped out-of-fold logistic evaluation keeps related answers together and excludes post-onset fitting rows. An offline HTML report displays risk traces, onset alignment, token records and audit metadata.
What’s implemented
- Annotation-aware pre-token feature extraction
- Grouped out-of-fold evaluation and censored-tail handling
- CSV features/predictions, metric records and a saved logistic model
- Offline interactive reports with state traces and timing audits
PROJECT STATUS / EXPLORATORY PROTOTYPE
Where it stands
Exploratory model-state prototype
- The shipped demo uses a tiny randomly initialized model and eight handcrafted annotations; it demonstrates plumbing, not predictive evidence
- Results depend on annotation quality, model, prompt format and numerical precision
- The advertised 8 GB CUDA setup was not measured in the repository's documented demonstration
Source & resources
Documentation behind this project page
Reviewed October 1, 2026. Project descriptions reflect a source review; repository validation claims were not independently reproduced.