fix: require sampling_logprob per content[] entry, no fallback

Always read the per-token logprob from content[].sampling_logprob (the exact
value the sampler drew with). Fail loud if any content[] entry lacks
sampling_logprob instead of substituting the rounded content[].logprob or 0.0,
so silently degraded logprobs never reach RL training. Removes the now-unused
_extract_entry_logprob helper.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: Dranoxgithub

eval_protocol/integrations/fireworks_v1_completions_client.py

@@ -83,19 +83,6 @@ def _normalize_token_id_sequence(values: Any) -> List[int]:
     return [int(x) for x in list(values)]
 
 
-def _extract_entry_logprob(entry: Dict[str, Any]) -> float:
-    """Return the per-token logprob from a ``content[]`` logprobs entry.
-
-    Prefer ``sampling_logprob`` (the exact, full-precision value the sampler
-    actually drew with) over ``logprob`` (a rounded/verification value). The
-    sampler value is what RL training needs for accurate inference KLD.
-    """
-    value = entry.get("sampling_logprob")
-    if value is None:
-        value = entry.get("logprob", 0.0)
-    return float(value) if value is not None else 0.0
-
-
 def _coerce_message_content_to_text(content: Any) -> str:
     if content is None:
         return ""
@@ -437,8 +424,15 @@ async def create_completion_from_prompt_ids(
                     f"at index {index}; cannot align per-token logprobs to token ids "
                     "without re-encoding. Refusing to return corrupted data."
                 )
+            if entry.get("sampling_logprob") is None:
+                raise RuntimeError(
+                    "Fireworks /v1/completions content[] entry is missing "
+                    f"sampling_logprob at index {index}. The sampling logprob is the "
+                    "exact value the sampler drew with and is required for correct "
+                    "inference KLD; refusing to substitute the rounded logprob or 0.0."
+                )
             completion_token_ids.append(int(entry["token_id"]))
-            completion_logprobs.append(_extract_entry_logprob(entry))
+            completion_logprobs.append(float(entry["sampling_logprob"]))
 
         completion_text = self.decode_token_ids(token_ids=completion_token_ids)
         if not completion_text:

tests/test_fireworks_v1_completions_client.py

@@ -174,7 +174,7 @@ def _fail_tokenizer():
                         "content": [
                             {"token_id": 271, "sampling_logprob": -0.05483185, "logprob": -0.0548313},
                             {"token_id": 248068, "sampling_logprob": -0.0014, "logprob": -0.0014},
-                            {"token_id": 26108, "logprob": -1.0},
+                            {"token_id": 26108, "sampling_logprob": -1.0, "logprob": -0.99},
                         ]
                     },
                 }
@@ -189,6 +189,34 @@ def _fail_tokenizer():
     asyncio.run(client.close())
 
 
+def test_raises_when_content_entry_missing_sampling_logprob(monkeypatch):
+    client = FireworksV1CompletionsClient(
+        model_id="test-model",
+        tokenizer_name_or_path="Qwen/Qwen3-0.6B",
+    )
+    monkeypatch.setattr(client, "decode_token_ids", lambda token_ids: "text")
+    _install_fake_completion(
+        client,
+        monkeypatch,
+        {
+            "choices": [
+                {
+                    "finish_reason": "stop",
+                    "logprobs": {
+                        "content": [
+                            {"token_id": 1, "sampling_logprob": -0.1},
+                            {"token_id": 2, "logprob": -0.2},
+                        ]
+                    },
+                }
+            ],
+        },
+    )
+    with pytest.raises(RuntimeError, match="missing .*sampling_logprob"):
+        asyncio.run(client.create_completion_from_prompt_ids(prompt_token_ids=[1]))
+    asyncio.run(client.close())
+
+
 def test_raises_when_no_content_logprobs(monkeypatch):
     client = FireworksV1CompletionsClient(
         model_id="test-model",