[MrScripty]

Implements docs/plans/v080-model-categories/plan.md for 0.8.0.

Models were grouped by one category picked from Hugging Face task tags and weak file and name hints. That category also decided the model's storage folder and ID. On the owner's library this filed Kev and Laya under LLM, LTX-2.3 under audio and embedding, and Granite Speech nar under embedding.

What changes

• Categories describe what a model is used for.
  • New decision and video categories.
  • A GGUF decision head (<arch>.decision.type, as in Kev and Laya) marks a decision model.
  • Video tasks map to video.
  • ...ForASR architectures map to audio.
  • The source's own task tag outranks derived hints.
  • A hard signal (architecture, config, decision head) is never outvoted by a task tag. When they disagree, confidence drops and the model is flagged for review.
• The category is metadata. Reclassification and migration keep the model's storage bucket (the first ID segment), so a category change never moves a folder or changes an ID. Only a family change still moves a model.
  • Name tokens no longer count the bucket as evidence. The bucket is only the last fallback.
  • ModelType::ALL is the single list of categories. Strict v2 metadata validation and the RPC derive from it.
• Inputs and outputs (modalities) are recorded for these tasks, including a new decision modality. When a strong source sets the category against a contradicting tag, the modalities follow the category. The tag stays recorded as evidence.
• A person's choice wins. set_model_category records a category (source user) that re-detection keeps, or clears it to re-detect. Requests are admitted exactly: a missing category, a task tag or unknown is refused with a reason.
• Desktop.
  • Each catalog row carries input and output modalities and an adapter flag. LoRA and PEFT packages are detected from their file names, adapter_config.json or the model name.
  • The library is grouped by use in a fixed order under readable headings.
  • Rows show "inputs → outputs" and an Adapter badge.
  • The metadata window has a "Used as" picker.
• Existing libraries. On open, system rule rows seeded by earlier versions are updated in place. Rules a person edited are left alone.

Verification

• Owner library (94 models), live refresh with the release build:
  • Kev and Laya → decision (text → decision).
  • LTX-2.3 ×3 → video (image → video).
  • Granite Speech nar → audio (audio → text).
  • The IC-LoRA is flagged as an adapter.
  • 0 reclassify errors, no model ID or folder changes, and a second pass is settled.
  • The first live run found strict-v2 validation rejecting the new categories (fixed in e078233; the tests now use strict v2 metadata and fail without the fix).
• Rust:
  • pumas-library: 2,363 passed.
  • pumas-rpc: 431 passed.
  • fmt and clippy (all features, -D warnings) are clean.
  • The Windows target check passes.
• Frontend: 949 tests pass, plus lint and types.
• Electron: all tests pass, along with lint, desktop-contract check and generator, and conformance in both the Electron and frontend suites.

Known limits

• The LTX IC-LoRA has no task tag or base-model evidence, so it stays under image generation until someone chooses Video.
• LTX's tag only states image → video.

🤖 Generated with Claude Code

[coderabbitai]

Important

• Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: ea59a295-68ff-4c69-a02a-148111790696

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The PR changes model classification and library organization across the shared library and desktop catalog.

Fix the category-clear path before merging; clearing a metadata choice can move a model and leave the desktop using its old ID.

Findings

1.  Clearing categories can move models ▶
2.  Decision outputs stay wrong ▶
3.  Conflicting tags go unflagged ▶

Summary

This PR makes model categories independent of storage folders and adds Decision and Video.

• Models are classified by their use, including decision and video.
• Category changes leave each model in its existing folder with the same ID.
• People can choose a model category and keep it through later scans.
• The library groups models by use and shows modalities and adapter badges.
• New audit and planning documents describe repository findings and future work.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Choose or clear category] --> B{Clear choice?}
  B -->|No| C[Save user category]
  B -->|Yes| D[Remove user source]
  D --> E[Reclassify category and family]
  E --> F{Family changed?}
  F -->|Yes| G[Move model within storage bucket]
  F -->|No| H[Keep folder and ID]
  G --> I[Read original model location]
  H --> I
  C --> J[Reindex model]
  J --> K[Project task and modalities]

Loading

_{Reviews (1) · Last reviewed commit: "docs(plan): record the owner-library acc..." · Reviewed by Greptile}

[greptile-apps]

Clearing categories can move models

When a model’s family is stale, clearing its chosen category calls reclassify_model, which can move its folder and change its ID. This code discards the returned ID and reads the old location. The request then fails or returns unknown, while the desktop keeps the old ID. Use an in-place detection path here so clearing a category keeps the promised folder and ID.

[MrScripty]

Fixed in 530d0fe: clearing a choice now re-detects the category only, in one guarded write, and never re-detects the family, so the model keeps its folder and ID. Test: clearing_a_choice_keeps_the_model_where_it_is.

[greptile-apps]

Decision outputs stay wrong

A decision GGUF without a task tag never gets decision outputs. apply_task_projection_from_persisted_evidence returns before consulting the detected category, leaving existing text outputs or unknown values unchanged through later scans. The new chosen_category_survives_reclassification_until_cleared fixture already has this shape, but only checks the category. Use strong category evidence when no usable tag exists, and check the task and modalities in that test.

[MrScripty]

Fixed in 530d0fe: when there is no task tag and the recorded task contradicts a strongly sourced category, the task (and inputs/outputs) follow the category, so a tagless decision GGUF reports text -> decision. Models with no recorded task are left unchanged. Covered in chosen_category_survives_reclassification_until_cleared.

[greptile-apps]

Conflicting tags go unflagged

The decision-head branch returns before checking task hints. It always gives confidence 0.95 with no review_reasons, even when the GGUF is tagged text-generation or feature-extraction. Later task projection does not add the missing warning. Keep Decision as the result, but lower confidence and flag conflicting tags as the architecture/config path does.

[MrScripty]

Fixed in 530d0fe: the decision head still decides the category, but a task tag naming another category now adds model-type-hint-conflict and model-type-low-confidence (confidence 0.75). Test: decision_head_with_a_contradicting_tag_is_flagged_for_review.