Use T5Tokenizer directly for Vectara HHEM v2

AutoTokenizer can't resolve HHEMv2Config (custom class, not registered).
HHEM v2 is T5-small based — T5Tokenizer loads without auto-detection.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Dockerfile

@@ -13,9 +13,9 @@ RUN pip install --no-cache-dir -r requirements.txt
 # Pre-download models so first request isn't slow on HF Spaces
 RUN python -c "\
 from sentence_transformers import SentenceTransformer; \
-from transformers import AutoTokenizer, pipeline; \
+from transformers import T5Tokenizer, pipeline; \
 SentenceTransformer('all-MiniLM-L6-v2'); \
-tok = AutoTokenizer.from_pretrained('vectara/hallucination_evaluation_model', trust_remote_code=True); \
+tok = T5Tokenizer.from_pretrained('vectara/hallucination_evaluation_model'); \
 pipeline('text-classification', model='vectara/hallucination_evaluation_model', tokenizer=tok, trust_remote_code=True)"
 
 COPY knowledge/ ./knowledge/

backend/grader.py

@@ -16,7 +16,7 @@ from typing import Any
 
 from sentence_transformers import SentenceTransformer
 from sklearn.metrics.pairwise import cosine_similarity
-from transformers import AutoTokenizer, pipeline as hf_pipeline
+from transformers import T5Tokenizer, pipeline as hf_pipeline
 
 from config import EMBEDDER_MODEL
 from rosetta import check_terminology
@@ -41,7 +41,7 @@ def get_nli_model() -> Any:
     """Return the shared Vectara faithfulness pipeline, loading it on first call."""
     global _nli_model
     if _nli_model is None:
-        tokenizer = AutoTokenizer.from_pretrained(NLI_MODEL, trust_remote_code=True)
+        tokenizer = T5Tokenizer.from_pretrained(NLI_MODEL)
         _nli_model = hf_pipeline(
             "text-classification",
             model=NLI_MODEL,