Spaces:
Sleeping
Sleeping
Commit History
-
revert: go back to in-Space inference (free CPU on Space)
HF deprecated free serverless inference for fine-tuned models like ours ("Model not supported by provider hf-inference"). Going back to the prior pattern: load the models inside this Space and run forward passes on its CPU. - Quick Vibe: AutoModelForSequenceClassification on itsLu/mentalbert-v5-flat-8class. - Deep Dive: snapshot_download the model repo and import its handler.py via importlib, then call EndpointHandler.__call__ with {inputs, mode}. Re-uses the model author's calibration and cascade logic as-is so any future tweak to the handler picks up automatically. Both models load lazily on first request, locked per model so concurrent requests don't double-download. Container start is instant; first user request is slow (~30-60s while weights download); subsequent calls are fast. HF_TOKEN remains used for private/gated downloads.8bff115 -
refactor: thin HF Inference API proxy for v5 models
Replace local torch + 4-stage cascade orchestration with two single HF Inference API calls: - Quick Vibe → itsLu/mentalbert-v5-flat-8class (8-class flat) - Deep Dive → itsLu/mentalbert-v5-hierarchical-v3 (5-stage cascade) New: sensitive_mode request param forwards to Deep Dive as mode=safety|balanced. Directed Aggression maps to "unhinged" for the existing frontend display. 503 cold-start triggers one retry using HF's estimated_time. Deep Dive confidence is derived from stage_probs at the exit stage. Drop torch + transformers + scikit-learn + joblib + sentencepiece; add httpx. Drop COPY of saved_models from Dockerfile. Stop tracking saved_models so the Space image stays slim.98ebcd8
-
chore: suppress transformers warnings via env var, clean up app.py
bc2f0f9
-
initial deployment
e5067f3