[v1] Add support for ShareGPT format. (#9486)

src/llamafactory/v1/plugins/data_plugins/converter.py

@@ -15,11 +15,15 @@
 
 from typing import Callable, TypedDict
 
-from typing_extensions import NotRequired
+from typing_extensions import NotRequired, Required
 
+from ....extras import logging
 from ...extras.types import DPOSample, Sample, SFTSample
 
 
+logger = logging.get_logger(__name__)
+
+
 class AlpacaSample(TypedDict, total=False):
     system: NotRequired[str]
     instruction: NotRequired[str]
@@ -27,6 +31,21 @@ class AlpacaSample(TypedDict, total=False):
     output: NotRequired[str]
 
 
+ShareGPTMessage = TypedDict(
+    "ShareGPTMessage",
+    {
+        "from": Required[str],  # Role of the message sender (e.g., "human", "gpt", "system")
+        "value": Required[str],  # Content of the message
+    },
+)
+
+
+class ShareGPTSample(TypedDict, total=False):
+    """Type definition for raw ShareGPT sample."""
+
+    conversations: Required[list[ShareGPTMessage]]
+
+
 class PairSample(TypedDict, total=False):
     prompt: NotRequired[str]
     chosen: NotRequired[list[dict]]
@@ -48,6 +67,20 @@ def alpaca_converter(raw_sample: AlpacaSample) -> SFTSample:
             {"role": "system", "content": [{"type": "text", "value": raw_sample["system"]}], "loss_weight": 0.0}
         )
 
+    if "history" in raw_sample:
+        for idx, item in enumerate(raw_sample["history"]):
+            if len(item) != 2:
+                logger.warning_rank0(
+                    f"Warning: History item at index {idx} has invalid length (expected 2, got {len(item)}). Skipping."
+                )
+                continue
+
+            old_prompt, old_response = item
+            messages.append({"role": "user", "content": [{"type": "text", "value": old_prompt}], "loss_weight": 0.0})
+            messages.append(
+                {"role": "assistant", "content": [{"type": "text", "value": old_response}], "loss_weight": 1.0}
+            )
+
     if "instruction" in raw_sample or "input" in raw_sample:
         messages.append(
             {
@@ -67,6 +100,62 @@ def alpaca_converter(raw_sample: AlpacaSample) -> SFTSample:
     return {"messages": messages}
 
 
+def sharegpt_converter(raw_sample: ShareGPTSample) -> SFTSample:
+    """Converts a raw ShareGPT sample into a formatted SFT (Supervised Fine-Tuning) sample.
+
+    Retains only SFT-relevant scenarios and removes parity checks.
+
+    Args:
+        raw_sample (ShareGPTSample): A raw sample in ShareGPT format.
+
+    Returns:
+        dict: A dictionary containing the formatted 'messages' list for SFT training.
+              Returns an empty list if the input data is invalid.
+    """
+    tag_mapping = {
+        "human": "user",
+        "gpt": "assistant",
+        "observation": "observation",
+        "function_call": "function",
+    }
+    messages = raw_sample.get("conversations", [])
+    aligned_messages = []
+    system_content = ""
+
+    # Extract system message if present (typically the first message)
+    if messages and messages[0]["from"] == "system":
+        system_content = messages[0]["value"]
+        messages = messages[1:]
+
+    if system_content:
+        aligned_messages.append(
+            {"role": "system", "content": [{"type": "text", "value": system_content}], "loss_weight": 0.0}
+        )
+
+    has_invalid_role = False
+    for message in messages:
+        sender = message["from"]
+        # validate sender is in supported tags
+        if sender not in tag_mapping:
+            logger.warning_rank0(f"Unsupported role tag '{sender}' in message: {message}")
+            has_invalid_role = True
+            break
+
+        aligned_messages.append(
+            {
+                "role": tag_mapping[sender],
+                "content": [{"type": "text", "value": message["value"]}],
+                "loss_weight": 0.0 if sender in ("human", "observation") else 1.0,
+            }
+        )
+
+    if has_invalid_role:
+        logger.warning_rank0("Skipping invalid example due to unsupported role tags.")
+        return {"messages": []}
+
+    return {"messages": aligned_messages}
+
+
 def pair_converter(raw_sample: PairSample) -> DPOSample:
     """Convert Pair sample to standard DPO sample.
 
@@ -148,6 +237,7 @@ def pair_converter(raw_sample: PairSample) -> DPOSample:
 CONVERTERS = {
     "alpaca": alpaca_converter,
     "pair": pair_converter,
+    "sharegpt": sharegpt_converter,
 }