[posttrain] Add Nemotron agentic SFT datasets

.agents/skills/lint-review/SKILL.md

@@ -7,8 +7,8 @@ allowed-tools: Bash(./infra/pre-commit.py:*), Bash(gh pr comment:*), Bash(gh pr
 # Skill: Lint-catalog review on a PR
 
 Run the `infra/lint/` catalog review (`./infra/pre-commit.py --review`)
-over a pull request's branch diff and surface every finding — as `file:line` 
-inline review comments where the finding's line is available, and as a 
+over a pull request's branch diff and surface every finding — as `file:line`
+inline review comments where the finding's line is available, and as a
 single fallback comment for the rest.
 
 ## Your contract

experiments/posttrain/instruction_datasets.py

@@ -38,6 +38,10 @@
 23. open-thoughts/OpenThoughts3-1.2M  # Original OT3 dataset; smoltalk2 uses a slightly different version
 24. lm-provers/FineProofs-SFT
 25. lm-provers/FineProofs-SFT/proof-only
+26. nvidia/Nemotron-SFT-SWE-v3
+27. nvidia/Nemotron-Cascade-SFT-SWE
+28. nvidia/Nemotron-SFT-OpenCode-v1
+29. nvidia/Nemotron-SFT-CUDA-v1
 """
 
 import dataclasses
@@ -103,6 +107,14 @@
 ]
 
 NEMOTRON_V1_SPLITS = ["chat", "code", "math", "stem", "tool_calling"]
+NEMOTRON_OPENCODE_V1_SPLITS = [
+    "general",
+    "bash_only_tool",
+    "bash_only_tool_skills",
+    "question_tool",
+    "agent_skills",
+    "agent_skills_question_tool",
+]
 
 
 @dataclass(frozen=True)
@@ -156,6 +168,33 @@ def multi_turn_adapter(
     )
 
 
+def rich_multi_turn_adapter(
+    conversation_column: str = "messages",
+    role_key: str = "role",
+    user_value: str = "user",
+    assistant_value: str = "assistant",
+    system_value: str = "system",
+    tool_value: str = "tool",
+    content_key: str = "content",
+    metadata_remap: dict[str, str] | None = None,
+    replacements: dict[str, str] | None = None,
+    extra_metadata_fn=None,
+) -> TransformAdapter:
+    return TransformAdapter(
+        dataset_format=InputDatasetFormat.SINGLE_COLUMN_MULTI_TURN_RICH,
+        conversation_column=conversation_column,
+        role_key=role_key,
+        user_value=user_value,
+        assistant_value=assistant_value,
+        system_value=system_value,
+        tool_value=tool_value,
+        content_key=content_key,
+        metadata_remap=metadata_remap or {},
+        replacements=replacements,
+        extra_metadata_fn=extra_metadata_fn,
+    )
+
+
 def instruction_response_adapter(
     *,
     instruction_column: str,
@@ -248,6 +287,22 @@ def __call__(self, row: dict[str, Any]) -> dict[str, Any]:
 SYNTHETIC2_SFT_VERIFIED_REVISION = "fce247fe48af8ff9624fb51d1de63aa1b2332cef"
 SYNTHETIC2_SFT_VERIFIED_METADATA_COLUMNS = ["problem_id", "task_type", "reward"]
 
+NEMOTRON_SFT_SWE_V3_REVISION = "3f73de64c1fe928a8f538fe45ccc10c228cc4c6a"
+NEMOTRON_SFT_SWE_V3_METADATA_COLUMNS = ["uuid", "license"]
+NEMOTRON_CASCADE_SFT_SWE_REVISION = "b4ef73ff9bbcfc33d1ec7a48e53017f9ce7af7a3"
+NEMOTRON_CASCADE_SFT_SWE_METADATA_COLUMNS = ["category", "source", "generator", "thinking"]
+NEMOTRON_SFT_OPENCODE_V1_REVISION = "556d5237acff203f3e1a0be49428634c3606cda2"
+NEMOTRON_SFT_OPENCODE_V1_METADATA_COLUMNS = [
+    "question_category",
+    "complexity_level",
+    "uuid",
+    "enabled_tools",
+    "skills_path",
+    "hf_split",
+]
+NEMOTRON_SFT_CUDA_V1_REVISION = "1a06167a6e1e90d928094184173898cbb9bf42de"
+NEMOTRON_SFT_CUDA_V1_METADATA_COLUMNS = ["uuid", "license", "used_in"]
+
 FINEPROOFS_SFT_REVISION = "73661e6"
 FINEPROOFS_SFT_METADATA_COLUMNS = [
     "category",
@@ -430,6 +485,36 @@ def __call__(self, row: dict[str, Any]) -> dict[str, Any]:
         subsets=["default"],
         splits=["train"],
     ),
+    "nvidia/Nemotron-SFT-SWE-v3": InstructionDatasetConfig(
+        hf_dataset_id="nvidia/Nemotron-SFT-SWE-v3",
+        revision=NEMOTRON_SFT_SWE_V3_REVISION,
+        adapter=rich_multi_turn_adapter(),
+        metadata_columns=NEMOTRON_SFT_SWE_V3_METADATA_COLUMNS,
+        name="nvidia/Nemotron-SFT-SWE-v3",
+        subsets=["default"],
+        splits=["train"],
+        max_parallelism=16,
+    ),
+    "nvidia/Nemotron-Cascade-SFT-SWE": InstructionDatasetConfig(
+        hf_dataset_id="nvidia/Nemotron-Cascade-SFT-SWE",
+        revision=NEMOTRON_CASCADE_SFT_SWE_REVISION,
+        adapter=multi_turn_adapter(),
+        metadata_columns=NEMOTRON_CASCADE_SFT_SWE_METADATA_COLUMNS,
+        name="nvidia/Nemotron-Cascade-SFT-SWE",
+        subsets=["default"],
+        splits=["train"],
+        max_parallelism=16,
+    ),
+    "nvidia/Nemotron-SFT-CUDA-v1": InstructionDatasetConfig(
+        hf_dataset_id="nvidia/Nemotron-SFT-CUDA-v1",
+        revision=NEMOTRON_SFT_CUDA_V1_REVISION,
+        adapter=rich_multi_turn_adapter(metadata_remap={"tools": "tools"}),
+        metadata_columns=NEMOTRON_SFT_CUDA_V1_METADATA_COLUMNS,
+        name="nvidia/Nemotron-SFT-CUDA-v1",
+        subsets=["default"],
+        splits=["train"],
+        max_parallelism=8,
+    ),
     "sherryy/tulu-3-sft-personas-instruction-following-expanded": InstructionDatasetConfig(
         hf_dataset_id="sherryy/tulu-3-sft-personas-instruction-following-expanded",
         revision="79ab2c4",
@@ -595,6 +680,25 @@ def __call__(self, row: dict[str, Any]) -> dict[str, Any]:
         splits=[split_name],
     )
 
+for split_name in NEMOTRON_OPENCODE_V1_SPLITS:
+    dataset_key = f"nvidia/Nemotron-SFT-OpenCode-v1/{split_name}"
+    INSTRUCTION_DATASET_NAME_TO_CONFIG[dataset_key] = InstructionDatasetConfig(
+        name=dataset_key,
+        hf_dataset_id="nvidia/Nemotron-SFT-OpenCode-v1",
+        revision=NEMOTRON_SFT_OPENCODE_V1_REVISION,
+        adapter=rich_multi_turn_adapter(
+            metadata_remap={
+                "agent_prompt": "agent_prompt",
+                "metadata": "source_metadata",
+                "tools": "tools",
+            },
+        ),
+        metadata_columns=NEMOTRON_SFT_OPENCODE_V1_METADATA_COLUMNS,
+        subsets=["default"],
+        splits=[split_name],
+        max_parallelism=4,
+    )
+
 
 def get_directory_friendly_dataset_name(hf_dataset_id: str) -> str:
     dataset_name = hf_dataset_id.replace("/", "--")

lib/marin/src/marin/transform/conversation/adapters.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import dataclasses
+import json
 from collections.abc import Callable
 from dataclasses import dataclass, field
 from enum import Enum
@@ -50,11 +51,52 @@ class InputDatasetFormat(str, Enum):
     """
 
     SINGLE_COLUMN_MULTI_TURN = "messages"
+    SINGLE_COLUMN_MULTI_TURN_RICH = "messages_rich"
     INSTRUCTION_RESPONSE = "instruction_response"
     INSTRUCT_COLUMN_RESPONSE = "instruct_column_response"
     INSTRUCT_MSG_RESPONSE = "instruct_msg_response"
 
 
+RICH_MESSAGE_EXTRA_KEYS = (
+    "name",
+    "tool_call_id",
+    "reasoning_content",
+)
+
+
+def _parse_tool_calls(value: Any) -> list[dict[str, Any]] | None:
+    if value is None:
+        return None
+
+    parsed = value
+    if isinstance(value, str):
+        stripped = value.strip()
+        if not stripped:
+            return None
+        try:
+            parsed = json.loads(stripped)
+        except json.JSONDecodeError:
+            return None
+
+    if isinstance(parsed, dict):
+        parsed = [parsed]
+    if not isinstance(parsed, list) or not parsed:
+        return None
+    if not all(isinstance(item, dict) for item in parsed):
+        return None
+    return parsed
+
+
+def _raw_tool_calls(value: Any) -> str | None:
+    if not isinstance(value, str):
+        return None
+
+    stripped = value.strip()
+    if not stripped:
+        return None
+    return stripped
+
+
 @dataclass
 class TransformAdapter:
     dataset_format: InputDatasetFormat = InputDatasetFormat.INSTRUCTION_RESPONSE
@@ -130,6 +172,31 @@ def transform_conversation_to_openai_format(
                 role = role_to_openai_role[conv[self.role_key]]
                 messages.append(OpenAIChatMessage(role=role, content=conv[self.content_key]))
             return messages
+        elif self.dataset_format == InputDatasetFormat.SINGLE_COLUMN_MULTI_TURN_RICH:
+            messages = []
+            role_to_openai_role = {
+                self.user_value: "user",
+                self.assistant_value: "assistant",
+                self.system_value: "system",
+                self.tool_value: "tool",
+            }
+            conversation = row[self.conversation_column]
+            for conv in conversation:
+                role = role_to_openai_role[conv[self.role_key]]
+                message: dict[str, Any] = {"role": role, "content": conv[self.content_key]}
+                for key in RICH_MESSAGE_EXTRA_KEYS:
+                    value = conv.get(key)
+                    if value:
+                        message[key] = value
+
+                tool_calls = _parse_tool_calls(conv.get("tool_calls"))
+                if tool_calls is not None:
+                    message["tool_calls"] = tool_calls
+                elif raw_tool_calls := _raw_tool_calls(conv.get("tool_calls")):
+                    message["raw_tool_calls"] = raw_tool_calls
+
+                messages.append(OpenAIChatMessage(**message))
+            return messages
         elif self.dataset_format == InputDatasetFormat.INSTRUCT_COLUMN_RESPONSE:
             messages = []
             instruction = row[self.instruction_column]