Upload folder using huggingface_hub

poetic/interpretation.py

@@ -16,6 +16,11 @@ def build_interpret_contents(poem: str, correction: str | None = None) -> str:
 
 _FENCE_RE = re.compile(r"^```[a-zA-Z]*\s*|\s*```$", re.MULTILINE)
 
+_FIELDS_RE = re.compile(
+    r'"understandingZh"\s*:\s*"(?P<zh>.+?)"\s*,\s*"briefEn"\s*:\s*"(?P<en>.+?)\s*$',
+    re.DOTALL,
+)
+
 
 def parse_interpretation(text: str) -> Interpretation:
     """Parse the model's reply into both audience fields.
@@ -42,6 +47,15 @@ def parse_interpretation(text: str) -> Interpretation:
             break
         except (json.JSONDecodeError, ValueError):
             continue
+    if obj is None:
+        # Last resort: generation sometimes stops before the closing brace.
+        # Pull the two fields straight out of the (unescaped) fragment.
+        match = _FIELDS_RE.search(candidates[-1])
+        if match:
+            obj = {
+                "understandingZh": match.group("zh").strip(),
+                "briefEn": match.group("en").strip().rstrip('"}').strip(),
+            }
     if obj is None:
         raise ValueError(f"Failed to parse interpretation JSON: {text[:120]}")
 

poetic/providers/local_zerogpu.py

@@ -107,7 +107,7 @@ def _interpret_gpu(poem: str, correction: str | None) -> str:
         )
     inputs = _tokenizer(prompt_text, return_tensors="pt").to(device)
     output = _llm.generate(
-        **inputs, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.95
+        **inputs, max_new_tokens=2048, do_sample=True, temperature=0.7, top_p=0.95
     )
     return _tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)