Resolve refs to commits so nothing serves a stale model or split

load(repo, 'main') was lru_cached on the literal 'main', pinning whichever
commit was current at first use. Refs now resolve to a sha per call and the
caches key on the sha. Same for the test split, which grew 46 -> 51 rows in
two days. Reports name both commits, and a Timer refreshes the pickers.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

README.md

@@ -63,6 +63,21 @@ sélectionne le chemin SetFit, sinon le modèle est chargé comme un
 L'onglet **Évaluation** rejoue le split de test et affiche la matrice de
 confusion puis les cas un par un, erreurs d'abord.
 
+## Rien n'est jamais périmé
+
+`main` est résolu en commit à chaque appel, et les caches sont indexés sur le
+commit — pas sur le nom de la branche. Un `train.py --push` est donc pris en
+compte sans redémarrage : sans cela, `main` désignait pour toujours le modèle
+chargé au premier clic. Le split de test est résolu de la même façon, ce qui
+compte : il est passé de 46 à 51 cas en deux jours.
+
+Le score affiché nomme les deux commits (`modèle @ sha · données sha`), donc un
+chiffre se rattache toujours à ce qui l'a produit.
+
+La liste des dépôts et des révisions se recharge toutes les
+`REFRESH_SECONDS` secondes en gardant la sélection en cours, et `↻` force le
+rechargement.
+
 ## Structure
 
 ```

app/config.py

@@ -4,6 +4,7 @@ ORG = "bee2link"
 MODEL_SEARCH = "breakdown-risk-"
 DATASET_REPO = "bee2link/breakdown-risk"
 CALLER_TURNS = 3
+REFRESH_SECONDS = 30
 TOKEN = os.environ.get("HF_TOKEN")
 
 LABEL_NAMES = {"risk": "Risque de panne", "no_risk": "Pas de risque"}

app/evaluation.py

@@ -7,6 +7,7 @@ import numpy as np
 from datasets import load_dataset
 
 from .config import DATASET_REPO, TOKEN, display
+from .hub import resolve
 from .predictors import Predictor, load
 
 BATCH_SIZE = 8
@@ -14,9 +15,17 @@ BATCH_SIZE = 8
 Track = Callable[[Iterable], Iterable]
 
 
+class Split(NamedTuple):
+    ids: list[str]
+    texts: list[str]
+    gold: list[str]
+    sha: str
+
+
 class Report(NamedTuple):
     repo: str
-    revision: str
+    model_sha: str
+    dataset_sha: str
     hits: int
     total: int
     ms_per_case: float
@@ -24,11 +33,20 @@ class Report(NamedTuple):
     cases: list[list]
 
 
-@lru_cache(maxsize=1)
-def test_split() -> tuple[list[str], list[str], list[str]]:
-    data = load_dataset(DATASET_REPO, split="test", token=TOKEN)
+@lru_cache(maxsize=2)
+def at_sha(sha: str) -> Split:
+    data = load_dataset(DATASET_REPO, split="test", revision=sha, token=TOKEN)
     names = data.features["label"].names
-    return list(data["id"]), list(data["text"]), [names[i] for i in data["label"]]
+    return Split(
+        ids=list(data["id"]),
+        texts=list(data["text"]),
+        gold=[names[i] for i in data["label"]],
+        sha=sha,
+    )
+
+
+def test_split() -> Split:
+    return at_sha(resolve(DATASET_REPO, "main", repo_type="dataset"))
 
 
 def confusion(labels: list[str], gold: list[str], predicted: list[str]) -> list[list]:
@@ -77,20 +95,21 @@ def run(
     predict: Predictor | None = None,
     track: Track | None = None,
 ) -> Report:
-    predictor = predict if predict is not None else load(repo, revision)
-    ids, texts, gold = test_split()
+    predictor, model_sha = (predict, revision) if predict else load(repo, revision)
+    split = test_split()
 
     started = time.perf_counter()
-    labels, probabilities = classify_all(predictor, texts, track)
+    labels, probabilities = classify_all(predictor, split.texts, track)
     elapsed = time.perf_counter() - started
 
     predicted = [labels[int(row.argmax())] for row in probabilities]
     return Report(
         repo=repo,
-        revision=revision,
-        hits=sum(want == got for want, got in zip(gold, predicted)),
-        total=len(gold),
-        ms_per_case=elapsed * 1000 / len(texts),
-        confusion=confusion(labels, gold, predicted),
-        cases=cases(ids, texts, gold, predicted, probabilities),
+        model_sha=model_sha,
+        dataset_sha=split.sha,
+        hits=sum(want == got for want, got in zip(split.gold, predicted)),
+        total=len(split.gold),
+        ms_per_case=elapsed * 1000 / len(split.texts),
+        confusion=confusion(labels, split.gold, predicted),
+        cases=cases(split.ids, split.texts, split.gold, predicted, probabilities),
     )

app/handlers.py

@@ -22,15 +22,18 @@ def classify(repo: str, revision: str, transcript: str) -> tuple[dict[str, float
         return {}, "", ""
 
     started = time.perf_counter()
-    labels, probabilities = load(repo, revision)([text])
+    predictor, sha = load(repo, revision)
+    labels, probabilities = predictor([text])
     scores = {display(name): float(p) for name, p in zip(labels, probabilities[0])}
-    return scores, text, f"`{(time.perf_counter() - started) * 1000:.0f} ms`"
+    elapsed = (time.perf_counter() - started) * 1000
+    return scores, text, f"`{sha[:7]}` · `{elapsed:.0f} ms`"
 
 
 def summary(report: evaluation.Report) -> str:
     return (
         f"### {report.hits}/{report.total} — exactitude {report.hits / report.total:.1%}\n"
-        f"`{report.repo}` · `{report.revision}` · {report.ms_per_case:.0f} ms par cas"
+        f"`{report.repo}` @ `{report.model_sha[:7]}` · données `{report.dataset_sha[:7]}`"
+        f" · {report.ms_per_case:.0f} ms par cas"
     )
 
 
@@ -39,10 +42,10 @@ def evaluate(repo: str, revision: str, progress=gr.Progress()):
         return NO_MODEL, [], []
 
     progress(0, desc=LOADING)
-    predictor = load(repo, revision)
+    predictor, sha = load(repo, revision)
     track = partial(progress.tqdm, desc=SCORING)
 
-    report = evaluation.run(repo, revision, predict=predictor, track=track)
+    report = evaluation.run(repo, sha, predict=predictor, track=track)
     return summary(report), report.confusion, report.cases
 
 
@@ -51,6 +54,13 @@ def pick_revision(repo: str) -> gr.Dropdown:
     return gr.Dropdown(choices=choices, value=choices[0][1] if choices else None)
 
 
-def pick_model() -> gr.Dropdown:
-    choices = model_repos()
-    return gr.Dropdown(choices=choices, value=choices[0] if choices else None)
+def refresh(repo: str, revision: str) -> tuple[gr.Dropdown, gr.Dropdown]:
+    repos = model_repos()
+    chosen = repo if repo in repos else (repos[0] if repos else None)
+    choices = revisions(chosen) if chosen else []
+    shas = [sha for _, sha in choices]
+    keep = revision if revision in shas else (shas[0] if shas else None)
+    return (
+        gr.Dropdown(choices=repos, value=chosen),
+        gr.Dropdown(choices=choices, value=keep),
+    )

app/hub.py

@@ -1,3 +1,4 @@
+import re
 from pathlib import Path
 
 from huggingface_hub import HfApi, hf_hub_download
@@ -6,6 +7,15 @@ from .config import MODEL_SEARCH, ORG, TOKEN
 
 api = HfApi(token=TOKEN)
 
+SHA = re.compile(r"[0-9a-f]{40}")
+
+
+def resolve(repo: str, revision: str, repo_type: str = "model") -> str:
+    if revision and SHA.fullmatch(revision):
+        return revision
+    info = api.repo_info(repo, revision=revision or None, repo_type=repo_type, token=TOKEN)
+    return info.sha
+
 
 def model_repos() -> list[str]:
     return sorted(model.id for model in api.list_models(author=ORG, search=MODEL_SEARCH))

app/predictors.py

@@ -9,7 +9,7 @@ from sentence_transformers import SentenceTransformer
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
 from .config import TOKEN
-from .hub import download, has_file
+from .hub import download, has_file, resolve
 
 Predictor = Callable[[list[str]], tuple[list[str], np.ndarray]]
 
@@ -42,7 +42,12 @@ def automodel(repo: str, revision: str) -> Predictor:
 
 
 @lru_cache(maxsize=4)
-def load(repo: str, revision: str) -> Predictor:
-    if has_file(repo, "model_head.pkl", revision):
-        return setfit(repo, revision)
-    return automodel(repo, revision)
+def at_sha(repo: str, sha: str) -> Predictor:
+    if has_file(repo, "model_head.pkl", sha):
+        return setfit(repo, sha)
+    return automodel(repo, sha)
+
+
+def load(repo: str, revision: str) -> tuple[Predictor, str]:
+    sha = resolve(repo, revision)
+    return at_sha(repo, sha), sha

app/ui.py

@@ -2,8 +2,8 @@ from pathlib import Path
 
 import gradio as gr
 
-from .config import display
-from .handlers import EXAMPLES, classify, evaluate, pick_model, pick_revision
+from .config import REFRESH_SECONDS, display
+from .handlers import EXAMPLES, classify, evaluate, pick_revision, refresh
 from .hub import model_repos, revisions
 from .text import HEADER, PLACEHOLDER, TRANSCRIPT_INFO
 
@@ -78,7 +78,7 @@ def build() -> gr.Blocks:
         gr.Markdown(HEADER)
 
         with gr.Row():
-            model, revision, refresh = selectors()
+            model, revision, refresh_button = selectors()
         selection = [model, revision]
 
         with gr.Tab("Classer"):
@@ -87,8 +87,10 @@ def build() -> gr.Blocks:
         with gr.Tab("Évaluation"):
             evaluate_button, report, progress_target = evaluation_tab()
 
+        timer = gr.Timer(REFRESH_SECONDS)
+        timer.tick(refresh, selection, selection, show_progress="hidden")
+        refresh_button.click(refresh, selection, selection)
         model.change(pick_revision, model, revision)
-        refresh.click(pick_model, None, model)
         gr.on(
             [run.click, transcript.submit],
             classify,