Add support for multiple Wikimedia ML models (not just revertrisk)

[xenacode-art]

Currently, PR #45 implements revertrisk (language-agnostic revert risk prediction) as the only
ML model for auto-review decisions. However, Wikimedia provides many other ML
models that could be useful for
different wikis and languages.

This issue proposes making the ML model system generic and configurable, allowing wikis to
choose which model(s) work best for their language and use case.

Background (from @zache-fi )

"The thing is that without testing we do not know (or least i do not know) how well
different models are working and probaply their performance differs depending on language. Ie.
some model would be better in certain language and older ORES models are probaply pretty good
in languages where there was lot of manual labelled training data, but not so good in
others."

Current State (PR #45)

• ✅ Revertrisk API integration implemented
• ✅ revertrisk_threshold field in WikiConfiguration
• ✅ Hardcoded to use only revertrisk model
• ⚠️ Not flexible for other models

Proposed Changes

1. Database Schema

Add fields to WikiConfiguration to support multiple models:

class WikiConfiguration(models.Model):
    # ... existing fields ...

    # New fields for generic ML support
    ml_model_type = models.CharField(
        max_length=50,
        default='revertrisk',
        choices=[
            ('revertrisk', 'Revert Risk (language-agnostic)'),
            ('articlequality', 'Article Quality'),
            ('articletopic', 'Article Topic'),
            ('damaging', 'Damaging Edit Detection'),
            ('goodfaith', 'Good Faith'),
            # More models from https://meta.wikimedia.org/wiki/Machine_learning_models
        ],
        help_text="Which Wikimedia ML model to use for this wiki"
    )

    ml_model_threshold = models.FloatField(
        default=0.0,
        help_text="Threshold for ML model score (0.0-1.0). Set to 0.0 to disable."
    )

2. Refactor Backend

- Abstract the ML API call logic from _get_revertrisk_score() into a generic
_get_ml_score(model_type, revision) function
- Support different API endpoints and response formats for different models
- Make the autoreview check model-agnostic

3. Vue.js Configuration Panel

Add model selection dropdown alongside the threshold input:
- Dropdown to select which ML model to use
- Threshold input (already being added in PR #45)
- Description of what each model does
- Link to Wikimedia ML models documentation

4. Documentation

- Document available models and their use cases
- Explain which models work best for which languages
- Provide examples of good threshold values

Available Wikimedia ML Models

From https://meta.wikimedia.org/wiki/Machine_learning_models:

<img width="991" height="214" alt="Image" src="https://github.com/user-attachments/assets/17152c5a-d59e-45d4-89a8-ebb9321991e1" />


Benefits

1. Language flexibility: Wikis can choose models that work best for their language
2. Testing capability: Compare different models to see which performs better
3. Future-proof: Easy to add new models as Wikimedia releases them
4. Graceful fallback: Can configure a fallback model if primary model fails

Implementation Phases

Phase 1: Backend Refactoring

- Refactor _get_revertrisk_score() to be model-agnostic
- Add support for 2-3 additional models (e.g., damaging, goodfaith)
- Update database schema with migrations

Phase 2: UI Support

- Add model selection dropdown to Vue.js config panel
- Show model-specific help text and documentation
- Allow per-wiki model configuration

Phase 3: Testing & Documentation

- Test different models on different language wikis
- Document which models work best where
- Provide recommended configurations

Questions

1. Should we support multiple models simultaneously (e.g., check BOTH revertrisk AND
damaging)?
2. Should model thresholds be per-model or global?
3. Should we implement automatic model selection based on wiki language?
4. Should there be a fallback model if the primary model API fails?

Related

- Builds on: #18, PR #45
- Wikimedia ML Models: https://meta.wikimedia.org/wiki/Machine_learning_models
- API Documentation: https://api.wikimedia.org/wiki/Lift_Wing_API

---
@zache-fi  requested this as a separate enhancement. Happy to work on this or collaborate with     
anyone interested!

[sergyDwhiz]

@xenacode-art happy to collab.

[xenacode-art]

Hi @zache-fi and @ad-an-26,

I wanted to bring issue #90 to your attention as it relates to the recent merge of PR #88 (ORES ML
support).

Current Status:
PR #88 successfully added ORES integration with damaging and goodfaith models, which partially addresses
the multi-model support requested in issue #90. The ORES implementation provides:

• Configurable damaging/goodfaith thresholds
• Living person biography checks with stricter thresholds
• Full UI support for model configuration
• Caching via the ModelScores table

Remaining Gap:
Issue #90 originally requested support for additional Wikimedia ML models beyond what ORES provides,
specifically:

• revertrisk (language-agnostic revert risk via Lift Wing API)
• articlequality (article quality assessment)
• articletopic (topic classification)

My Question:
Given that PR #88 already provides multi-model support through ORES (damaging + goodfaith), should issue
#90 be:

1. Closed as resolved - Since we now have multiple ML models through ORES, which satisfies the core
   request for language/wiki-specific model flexibility?
2. Kept open with revised scope - To add support for non-ORES models (revertrisk, articlequality,
   articletopic) via Wikimedia's Lift Wing API as a future enhancement?
3. Converted to a research/testing task - To evaluate which models work best for different language
   wikis and document recommendations?

I'm happy to work on option 2 or 3 if you think there's value in expanding beyond ORES, but wanted to
get your guidance on the priority and approach given that the core multi-model functionality is now
available.

Thanks for your input!

[sergyDwhiz]

Hi @xenacode-art — thanks for working on this!

Issue:
The database fields were added, but the implementation looks incomplete. The autoreview.py script doesn’t yet use ml_model_type or ml_model_threshold — it still only checks ORES thresholds.

What’s left to do:

• Autoreview integration: Make autoreview.py read configuration.ml_model_type and route to the correct API.
• API views: Update api_configuration to handle the new model type field.
• Frontend: Add a dropdown to select which ML model to use.
• Tests: Add the test_ml_models.py file mentioned in the commit.

Why it matters:
Right now, ORES (from PR #88) works, but your PR would allow wikis to choose between ORES, LiftWing RevertRisk, and other models.

Next steps:
Would you like help wrapping up the remaining pieces? I could create a PR featuring the autoreview integration, which you could then pull from and merge.

[xenacode-art]

Thanks @sergyDwhiz for the review, I will push a commit that handles these shortly.

[xenacode-art]

Hi @sergyDwhiz,

Implementation complete! I've pushed the full Lift Wing ML model integration to the feature/18 branch.

What's included:
✅ Database fields (ml_model_type, ml_model_threshold)✅ Backend autoreview integration with routing to
correct API✅ API views updated to handle model configuration✅ Frontend dropdown with 5 model choices✅
Migration file✅ Full UI with validation and help text

The implementation now allows wikis to choose between 5 Lift Wing ML models (revertrisk, damaging,
goodfaith, articlequality, articletopic) in addition to the existing ORES models. Both systems work
independently and can be configured separately.

Ready for review by @zache-fi and @ad-an-26

[sergyDwhiz]

@xenacode-art, this is good. All looks good by me.
Waiting for @zache-fi review. Hopefully it gets merged!

Twas nice collaborating..