[Draft] Show how each model performs, and let people start the jobs behind it

[mohamedelabbas1996]

Summary

Antenna can now score a model against a fixed set of verified occurrences, and retrain a classifier head from labels people confirmed. Neither was visible: the numbers existed only in the API, and the jobs behind them could only be started by posting to it by hand. This is the interface for both.

Split out of #1407 so the screens can be reviewed on their own. The API fields they read land in that PR; until it merges the new columns read n/a rather than breaking.

Screenshots

Taxa list — verified crops ready to train a head on, per species	Table columns — the column can be switched off

A species that has been scored — every model, its accuracy here, and the set used	A species nothing has scored yet — reads n/a, not 0

Taxa lists — the model that handles the list's species best, and the average it won on	Algorithm panel — training data ready, and what the model has been scored on

Jobs table — the new job types are named instead of blank	The same table filtered to Evaluate algorithm

Create job — defaults to an ML pipeline run, as before	Train classifier — asks for a head, not captures

Evaluate algorithm — asks for a head and an evaluation set	Generate embeddings — asks for a pipeline

The head picker offers only heads that can be retrained

List of Changes

#	What it does for a user	How
1	A species list shows the verified crops ready to train a head on	training_crops_ready, requested only when the column is visible
2	A species page lists every model scored on it, with that species' accuracy and the set used	algorithm_performance on the taxon, each row linking to the model
3	A taxa list names the model that handles its species best	best_model, with the per-species average it was ranked on
4	An algorithm's panel shows what it has been scored on	evaluations table in the details dialog
5	It also shows whether a retrain is worth starting	crops ready, species covered, and crops still waiting for an embedding
6	Training, embedding and evaluation jobs can be started from the jobs form	job type picker; each type asks only for the fields it needs
7	Those jobs show their type in the table and can be filtered by it	they were blank before, because the frontend's job type list had never been extended

Detailed Description

• Null is not zero. Every new field reads n/a when the API returns null — on a project where nothing has been evaluated, and on any deployment where Retrain a classifier head from verified identifications, and score what it produces #1407 has not landed. The crop count is deliberately null rather than 0 when it was not requested; a zero would read as a species with nothing to train on.
• The form reads each job type's declared requirements: a pipeline to process or embed, a head to retrain, a head and an evaluation set to score. It no longer asks for captures for a job that never looks at one.
• The crop count is opt-in. ?with_training_crop_counts=true is sent only when that column is visible, following the pattern Add example occurrence to the taxa list to expedite the verification of species presence #1365 established for the Example column. The count grows with the project rather than the page, and most viewers never show it.

How to Test the Changes

Against a project with verified occurrences and at least one scored algorithm:

1. Taxa — the "Training images ready" column, and the same count under Table columns.
2. Taxa › a species — "Algorithms & performance", one line per scored model, each opening that model.
3. Taxa lists — "Best model", naming the model and the per-species average it won on.
4. Algorithms › one algorithm — "Training data" and "Evaluation" sections.
5. Jobs › Create new job — switch Type and watch the fields change; create an Evaluate algorithm job and start it.

Verified in a browser against a local stack: a job created from this form ran and scored 1.000 over 15 occurrences. yarn type-check, yarn lint, yarn format --check and yarn test (52 tests) all pass.

Known gaps

1. The job form offers types the API rejects until Retrain a classifier head from verified identifications, and score what it produces #1407 lands. Picking Train classifier before then returns a validation error rather than a job.
2. The taxa list shows the best model's per-species average only; the plain share comes back in the same payload and is not displayed.

🤖 Generated with Claude Code

[netlify]

👷 Deploy request for antenna-ssec pending review.

Visit the deploys page to approve it

Name	Link
🔨 Latest commit	6ab0ebe

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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

_{Comment @coderabbitai help to get the list of available commands.}

[netlify]

✅ Deploy Preview for antenna-preview ready!

Name	Link
🔨 Latest commit	6ab0ebe
🔍 Latest deploy log	https://app.netlify.com/projects/antenna-preview/deploys/6aac541cdad41e0008eaf60f
😎 Deploy Preview	https://deploy-preview-1423--antenna-preview.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
Lighthouse	1 paths audited Performance: 56 (🔴 down 9 from production) Accessibility: 81 (🔴 down 8 from production) Best Practices: 92 (🔴 down 8 from production) SEO: 92 (no change from production) PWA: 80 (no change from production) View the detailed breakdown and full score reports
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[mihow]

Claude says: Hi! A heads-up rather than a request. Michael asked for one Create Job dialog that every job type can appear in, and a draft of it is up in #1447. It adds GET /api/v2/jobs/types/?project_id=N, which lists the job types a member may create in a project, what each one runs on, and a JSON Schema of its settings. The dialog builds its form from that response.

It is written to fit alongside #1407 and this PR rather than replace anything in them:

• required_fields, required_params and user_creatable on JobType use the same names and defaults as Retrain a classifier head from verified identifications, and score what it produces #1407, and the endpoint returns them for each type. So train_classifier, generate_embeddings and evaluate_algorithm would show up in the dialog once both land, with a simple field for each required param.
• If one of those jobs later gets a pydantic config_schema, it gets a generated form (labels, help text, pickers, numeric bounds) with no frontend change.
• The job-type picker here could read its list from the endpoint instead of a hardcoded one, if that turns out to be useful. No pressure either way.

Nothing needs to change in this PR for now. Happy to line things up whenever it suits you.

[mihow]

Claude says: A correction to my earlier note. After review, #1447 no longer uses required_fields / required_params. Each job type the Create Job dialog offers now declares one pydantic model for everything a new job takes, and a new job's inputs all arrive in params.config (with job_type_key required). So for train_classifier, generate_embeddings and evaluate_algorithm to appear in the dialog, each would declare a small model, for example an algorithm_key field, instead of the tuples. The dialog then generates the form from it. The job-type list on GET /jobs/types/ is otherwise as described. Sorry for the churn, and happy to help line it up when you get there.

[mohamedelabbas1996]

Moved to the lab repo and rebased onto the evaluation API it reads: #1506.

This branch lived on my fork, targeted main, and carried a copy of the training job form from before the retraining work was split. That form was rebuilt in #1504, and the training-data tab it added needs an endpoint from #1502, so neither came across. #1506 is the evaluation UI only, sitting on #1493.

Closing this one.