qev / ROADMAP.md
ken-jo's picture
Release QEV 0.1.1: Qwen3.5-2B typed decisions
f996835 verified
|
Raw History Blame Contribute Delete
2.26 kB

Research roadmap

Compact model size requirement

The next architecture targets 300-450M total parameters across vision, text and the decision head. RLCD is a probability-learning experiment, not a model-compression method. See the compact-model plan for design, distillation and evaluation gates. These compact weights do not exist in the current QEV release.

Closed Qwen-based line: QEV 0.1.0

Freeze the evaluated Qwen3.5-2B adaptation, publish the custom runtime, licensed dataset snapshots, exact evidence and known failures. Future bug fixes must preserve or explicitly revalidate the frozen inference contract. Do not market this checkpoint as a new base network.

Separate next network

After this publication is complete, begin a distinct experiment with an original integrated decision network. Its name and results will be separate from this checkpoint.

  1. Establish independent perception tests: local object identity, digits, positions, relations and conflicting evidence. Reserve unseen renderers, objects, compositions and photo domains.
  2. Prototype image tokens retaining spatial coordinates, text/criterion encoders, candidate- conditioned evidence attention and one shared candidate scorer. Include typed ordinal and binary outputs; test candidate permutation and paraphrase invariance.
  3. Compare against the frozen QEV baseline at equal data and hardware budgets. Use Qwen representations or teachers only where a measured benefit justifies them; backbone choice is open. Separate perception, evidence binding, decision logic and probability errors.
  4. Study supervised proper scoring, distillation and RL only after perception works. Select on development data, fit calibration separately and evaluate once on an untouched final set.
  5. Target resident latency below 300 ms on RTX 4060 Ti, with a declared image/token/question workload. Report p50/p95, warm/cold conditions and throughput separately.

This is a research plan, not a claim that the proposed network already exists or meets these goals. LAYA's interface and candidate scoring are references; proprietary JEV internals are unknown. Success requires measurable held-out transfer, not an impressive single game demo.