Spaces:

---

karlexmarin

/

taf-agent

like 0

Running

🔬 TAF Agent v0.10 — "what does your model's context length actually mean?" (free, browser-only)

The web is full of new long-context models — Mamba, RWKV, DeltaNet, Jamba, Zamba, Nemotron-H — all advertising big context numbers. But a "1M context" means something very different for a state-space model than for full attention. v0.10 adds three tools to cut through that, all running 100% in your browser (no inference, no signup):

• 🧠 Memory Reality Check — paste a model id → it detects the architecture (full-attention / sliding-window / SSM-Mamba / RWKV / linear / test-time-training / hybrid) straight from config.json and tells you what its "context length" really means and how it fails. e.g. state-space models read the whole stream but compress it into a fixed-size state — so exact needle-recall of an old token fails. Detection validated against 13 live HF configs.
• 📊 Prediction vs Reality — don't take the numbers on faith. Compare the tool's closed-form predictions against MEASURED values (the shipped dataset, or the JSON from the Diagnose CLI on your own weights), with a confidence score. And contribute your measurement back to the public dataset via a one-click PR — server-less, so it benefits everyone.
• ✅ Confidence score on every verdict — a 0–100% rating with a ✓/⚠ evidence checklist (γ measured vs closed-form, validated regime, benchmark available). Predictions are never presented as absolute truth.

Also in v0.10: HF model-id autocomplete on every input, a manual reorganised into scannable cards, and an honest fix to the long-context verdict (we found and disclosed a tautology in our own closed-form horizon — d_horizon ≡ T_eval — see docs/LIMITATIONS.md).

27 modes total, 4 languages (EN/ES/FR/ZH), 37 Lean+Mathlib-verified identities, 0 telemetry.

Try it: https://huggingface.co/spaces/karlexmarin/taf-agent
Feedback welcome — especially if a prediction disagrees with your real measurement (that's exactly what Prediction-vs-Reality is for). Paper: https://zenodo.org/records/20314038