Models for CritiQ v2, a seven-tier learned disk I/O prioritisation system
(final-year major project, SIT Tumakuru, ISE, 2025–26). They are trained in Google
Colab from the private dataset repo CipherSmit/critiq-v2-data and served by the CritiQ Space.
Folder
Model
Role (tier)
Format
seq/
Trace Transformer: causal Transformer over request tokens, pre-trained to predict the next request
Tier 2: actor class from its recent requests
safetensors
rf/
Random Forest (scikit-learn) on 12 hand-made features
baseline classifier (ablation only)
skops
vae/
beta-VAE, 60→…→32 latent + aux P99 head
Tier 3: workload fingerprint, OOD scores
safetensors + JSON
vae_plain/
same architecture, β = 0, no aux loss
ablation only
safetensors + JSON
mlp/
Policy MLP 38→64→32→4
Tier 5: priority P0–P3 per actor
safetensors
Training is sequential (classifier → beta-VAE → MLP): the VAE's input includes the classifier's
class mix, and the MLP's input includes the class, its confidence and the VAE fingerprint.
Architectures and feature order live in critiq-colab/critiq_models/ of the project repository.
Intended use and limits
Research use in trace-driven simulation (MSR Cambridge, FIU, Alibaba block traces).
Not validated on live systems. Class labels come from transparent behaviour rules over
actor features (or a hand-built FIU process→class map); priority labels come from
counterfactual simulation of an approximate mq-deadline scheduler with priority classes. Results depend on the simulator's fidelity (see the
calibration report in the project's research/ folder).
