Appearance

HLM3-Mix Model Lab ​

Browser UI for prompt-testing packaged HLM3 language checkpoints with explicit load/unload, token-level output, and per-layer iteration signals.

Status

Demo-ready browser surface. Wraps the language checkpoint catalog that ships with the on-site demo bundle; the same catalog also drives the all-models editions of the Critical Document Audit and AI Audit Showcase.

What the demo proves ​

A reviewer can directly compare HLM3-Mix checkpoints without juggling multiple terminals or duplicate model weights:

  1. Select a checkpoint from the sidebar — each option is labelled with its tier badge and the relevant benchmark metric.
  2. Load explicitly so only one model occupies VRAM at a time.
  3. Generate with greedy or sampling settings and inspect the per-token table.
  4. Unload before switching to the next checkpoint.
  5. Compare any two checkpoints sequentially with a single click.

What the reviewer sees ​

  • Live sidebar caveat and tier badge for the selected checkpoint
  • Token table with step, token_id, decoded token, and top-1 probability
  • Per-layer average Hopfield iterations for the run
  • A catalog table at the bottom listing every available entry plus on-disk status
  • Tokens-per-second, total tokens, and total seconds for each run

Why one catalog matters ​

The lab and the all-models editions of the audit UIs share a single source of truth. Adding a new checkpoint means editing one file; every downstream UI updates automatically.

Caveats ​

  • Some checkpoints in the bundle are diagnostic surfaces, not quality claims — the sidebar caveat and tier badge make the distinction explicit so a reviewer cannot mistake a scale or research artifact for a validated result.
  • Benchmark values come from offline evaluations; treat them as relative scale indicators, not benchmark headlines.
  • Larger checkpoints can need several GB of VRAM. The lab promotes large models to bfloat16 on CUDA automatically, but CPU mode is the safe fallback.

Where it fits ​

The right demo for:

  • Engineering reviewers comparing checkpoint behavior side-by-side
  • Partners evaluating which checkpoint to baseline against
  • Anyone who wants to see the validated public result reproduced live