Upload 89 files

app/engine/paraphrase/__init__.py

@@ -173,40 +173,49 @@ def _generate_raw(
     tokenizer, model = loaded
     import torch
 
-    prompt = f"paraphrase: {text.strip()}"
-    encoded = tokenizer(
-        prompt,
-        return_tensors="pt",
-        truncation=True,
-        max_length=256,
-    )
+    prompts = [
+        f"paraphrase: {text.strip()}",
+        f"rewrite with different wording but same meaning: {text.strip()}",
+    ]
     returns = max(1, num_return)
-    groups = min(returns, 4)
-    beams = max(6, returns * 2)
-    beams = max(groups, (beams // groups) * groups)
-    with torch.no_grad():
-        try:
-            outputs = model.generate(
-                **encoded,
-                max_new_tokens=ENGINE_PARAPHRASE_MAX_NEW_TOKENS,
-                num_beams=beams,
-                num_beam_groups=groups,
-                diversity_penalty=1.0,
-                num_return_sequences=returns,
-                do_sample=False,
-                early_stopping=True,
-            )
-        except Exception:
-            # Fall back to plain beam search if diverse beams are unsupported.
-            outputs = model.generate(
-                **encoded,
-                max_new_tokens=ENGINE_PARAPHRASE_MAX_NEW_TOKENS,
-                num_beams=max(4, returns),
-                num_return_sequences=returns,
-                do_sample=False,
-                early_stopping=True,
-            )
-    decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)
+    per_prompt = max(1, returns // len(prompts))
+    if per_prompt * len(prompts) < returns:
+        per_prompt += 1
+    decoded: list[str] = []
+    for prompt in prompts:
+        encoded = tokenizer(
+            prompt,
+            return_tensors="pt",
+            truncation=True,
+            max_length=256,
+        )
+        prompt_returns = per_prompt
+        groups = min(prompt_returns, 4)
+        beams = max(6, prompt_returns * 2)
+        beams = max(groups, (beams // groups) * groups)
+        with torch.no_grad():
+            try:
+                outputs = model.generate(
+                    **encoded,
+                    max_new_tokens=ENGINE_PARAPHRASE_MAX_NEW_TOKENS,
+                    num_beams=beams,
+                    num_beam_groups=groups,
+                    diversity_penalty=1.0,
+                    num_return_sequences=prompt_returns,
+                    do_sample=False,
+                    early_stopping=True,
+                )
+            except Exception:
+                # Fall back to plain beam search if diverse beams are unsupported.
+                outputs = model.generate(
+                    **encoded,
+                    max_new_tokens=ENGINE_PARAPHRASE_MAX_NEW_TOKENS,
+                    num_beams=max(4, prompt_returns),
+                    num_return_sequences=prompt_returns,
+                    do_sample=False,
+                    early_stopping=True,
+                )
+        decoded.extend(tokenizer.batch_decode(outputs, skip_special_tokens=True))
     cleaned: list[str] = []
     seen: set[str] = set()
     source_key = text.strip().lower().rstrip(".!?")