params changed

app.py

@@ -177,7 +177,7 @@ def _generate_batch(prompts: list[str]) -> list[str]:
     with torch.no_grad():
         out = _model.generate(
             **enc,
-            max_new_tokens=2048,   # ≈ 200 words — enough for our tight prompt
+            max_new_tokens=4096,   # ≈ 200 words — enough for our tight prompt
             use_cache=True,
             do_sample=False,
             pad_token_id=_tokenizer.pad_token_id,
@@ -202,9 +202,11 @@ def _generate_sequential(prompts: list[str]) -> list[str]:
         with torch.no_grad():
             out = _model.generate(
                 **enc,
-                max_new_tokens=1024,
+                max_new_tokens=16000,
                 use_cache=True,
                 do_sample=False,
+                temperature=0.4,
+                top_p=0.95,
                 pad_token_id=_tokenizer.pad_token_id,
             )
         text = _tokenizer.decode(out[0][enc.input_ids.shape[1]:], skip_special_tokens=True)