feat: rlm renderer for the RLM chat format

renderers/__init__.py

@@ -7,10 +7,10 @@
     __version__ = "0+unknown"
 
 from renderers.base import (
+    MULTIMODAL_MODELS,
     Content,
     ContentPart,
     ImagePart,
-    MULTIMODAL_MODELS,
     Message,
     MultiModalData,
     MultimodalRenderer,
@@ -43,29 +43,30 @@
 from renderers.configs import (
     AutoRendererConfig,
     BaseRendererConfig,
-    config_from_name,
-    DefaultRendererConfig,
     DeepSeekR1RendererConfig,
     DeepSeekV3RendererConfig,
+    DefaultRendererConfig,
+    GLM5RendererConfig,
     GLM45RendererConfig,
     GLM51RendererConfig,
-    GLM5RendererConfig,
     GptOssRendererConfig,
     Hy3RendererConfig,
-    KimiK25RendererConfig,
     KimiK2RendererConfig,
+    KimiK25RendererConfig,
     LagunaXS2RendererConfig,
     LagunaXS21RendererConfig,
     Llama3RendererConfig,
     MiniMaxM2RendererConfig,
     Nemotron3RendererConfig,
     Nemotron3UltraRendererConfig,
     PrimeQwen3RendererConfig,
-    Qwen35RendererConfig,
-    Qwen36RendererConfig,
     Qwen3RendererConfig,
     Qwen3VLRendererConfig,
+    Qwen35RendererConfig,
+    Qwen36RendererConfig,
     RendererConfig,
+    RlmRendererConfig,
+    config_from_name,
 )
 
 # Concrete renderer classes are lazy-loaded so that consumers needing
@@ -101,6 +102,7 @@
     "Qwen36Renderer": "renderers.qwen36",
     "Qwen3Renderer": "renderers.qwen3",
     "Qwen3VLRenderer": "renderers.qwen3_vl",
+    "RlmRenderer": "renderers.rlm",
 }
 
 
@@ -175,6 +177,8 @@ def __dir__() -> list[str]:
     "Qwen3RendererConfig",
     "Qwen3VLRenderer",
     "Qwen3VLRendererConfig",
+    "RlmRenderer",
+    "RlmRendererConfig",
     "RenderedConversation",
     "RenderedTokens",
     "RenderedTrainingSample",

renderers/base.py

@@ -1066,6 +1066,8 @@ def bridge_to_next_turn(self, *args: Any, **kwargs: Any) -> "RenderedTokens | No
     "meta-llama/Llama-3.2-3B-Instruct": "llama-3",
     # Poolside Laguna. The two checkpoints ship different chat templates,
     # each mirrored by its own renderer class.
+    # RLM chat format (minimal role tags, rlm harness).
+    "PrimeIntellect/RLM-Chat-Template": "rlm",
     "poolside/Laguna-XS.2": "laguna-xs.2",
     "poolside/Laguna-XS-2.1": "laguna-xs-2.1",
     # GPT-OSS.
@@ -1324,6 +1326,7 @@ def _populate_registry():
     from renderers.qwen3_vl import Qwen3VLRenderer
     from renderers.qwen35 import Qwen35Renderer
     from renderers.qwen36 import Qwen36Renderer
+    from renderers.rlm import RlmRenderer
 
     RENDERER_REGISTRY.update(
         {
@@ -1347,6 +1350,7 @@ def _populate_registry():
             "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "nemotron-3-ultra": Nemotron3UltraRenderer,
+            "rlm": RlmRenderer,
             "gpt-oss": GptOssRenderer,
         }
     )

renderers/configs.py

@@ -663,6 +663,29 @@ class DeepSeekR1RendererConfig(BaseRendererConfig):
     name: Literal["deepseek-r1"] = "deepseek-r1"
 
 
+class RlmRendererConfig(BaseRendererConfig):
+    """RLM chat format (``PrimeIntellect/RLM-Chat-Template``) — minimal
+    role-tag-only format for the rlm harness.
+
+    One message, one tag block: ``<system> <user> <assistant> <output>``
+    plus the single inline tool call ``<ipython>...</ipython>``. Thinking is
+    structurally never dropped, so ``thinking_retention`` may only be left
+    unset or ``"all"``. Tool schemas are never rendered; ``tools=`` is
+    validation-only (exactly ``[ipython]`` is accepted).
+    """
+
+    name: Literal["rlm"] = "rlm"
+
+    @model_validator(mode="after")
+    def _check_thinking_retention(self):
+        if self.thinking_retention not in (None, "all"):
+            raise ValueError(
+                "rlm never drops thinking; thinking_retention must be unset "
+                f"or 'all', got {self.thinking_retention!r}"
+            )
+        return self
+
+
 RendererConfig = Annotated[
     Union[
         AutoRendererConfig,
@@ -687,6 +710,7 @@ class DeepSeekR1RendererConfig(BaseRendererConfig):
         Nemotron3UltraRendererConfig,
         DeepSeekV3RendererConfig,
         DeepSeekR1RendererConfig,
+        RlmRendererConfig,
     ],
     Field(discriminator="name"),
 ]
@@ -727,6 +751,7 @@ class DeepSeekR1RendererConfig(BaseRendererConfig):
     "nemotron-3-ultra": Nemotron3UltraRendererConfig,
     "deepseek-v3": DeepSeekV3RendererConfig,
     "deepseek-r1": DeepSeekR1RendererConfig,
+    "rlm": RlmRendererConfig,
 }