vector-caliper: first real-data baseline (Qwen LoRA run) + dogfood notes#8
Closed
mcp-tool-shop wants to merge 1 commit into
Closed
vector-caliper: first real-data baseline (Qwen LoRA run) + dogfood notes#8mcp-tool-shop wants to merge 1 commit into
mcp-tool-shop wants to merge 1 commit into
Conversation
…dogfood notes 8-checkpoint trajectory from a production Qwen-Image style-LoRA run (RTX 5090), captured as ModelState series + rendered through the real pipeline. Headline: geometry (anisotropy 8.2->12.5, effdim collapse) caught checkpoint-2000 overfit that performance (CLIP-sim) alone rewarded — first live validation of the combined-view thesis. Dogfood notes: unseeded PCA breaks the determinism guarantee, dist is not Node-ESM-consumable (extensionless imports), README/package scope mismatch, required uncertainty fields need a diffusion-LoRA preset, and the canonical demo bypasses the public pipeline. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
Author
|
Withdrawn — prototypes is a seed vault; vector-caliper is being resurrected as a local working repo instead (baselines + dogfood notes land there). |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
First production training run fed through VectorCaliper end-to-end, per the "establish baselines, then predict/hypothesize" direction.
What this adds
baselines/qwen-lora-tallow-fen-v1.json— 8 checkpoint states from a real Qwen-Image style-LoRA run (rank 16, RTX 5090, 2000 steps): CLIP embedding-cloud geometry + performance per checkpoint, uncertainty via documented proxies.baselines/qwen-lora-tallow-fen-v1.svg— rendered through the actualProjectionEngine → SemanticMapper → SceneBuilder → SVGRendererpipeline (apparently its first real-data artifact; the canonical demo bypasses it).baselines/README.md+DOGFOOD-NOTES-2026-06-09.md.The headline
Between steps 1750→2000, performance.accuracy ROSE (CLIP-sim 0.7796→0.7937) while geometry collapsed (anisotropy 8.2→12.5, effective dimension 7.0→6.76). Performance-only selection picks the overfit checkpoint; geometry+performance picks step 1250 — which independently matched the CMMD minimum AND the human eyeball call (monochrome drift on neutral subjects). The combined view caught what the single metric missed. That is the tool thesis, validated on its first real run.
Dogfood findings (full notes in the file)
Math.random()— the seeded Mulberry32 in the same file is unused → determinism guarantee is not byte-level.tscemits extensionless/directory relative imports → raw Node ESM cannot load dist (ERR_UNSUPPORTED_DIR_IMPORT); 32 files patched locally to render. Fix:moduleResolution: NodeNext+.jsextensions in source.@mcp-tool-shop/vs package@mcptoolshop/(+private: true);repository.urlpoints at a different repo.uncertainty.*fields have no native analogue for diffusion-LoRA runs — suggest an optional group or a domain preset with blessed proxies.🤖 Generated with Claude Code