Formatters

# All formats via Dataset.save()
dataset.save("output.jsonl", format="messages")    # OpenAI (default)
dataset.save("output.jsonl", format="chatml")      # ChatML
dataset.save("output.jsonl", format="qa")          # Q&A
dataset.save("output.jsonl", format="langsmith")   # LangSmith
dataset.save("output.jsonl", format="langfuse")    # Langfuse
dataset.save("output.jsonl", format="tool_call")   # Tool calling
dataset.save("output.jsonl", format="bert")        # BERT classification
dataset.save("output.jsonl", format="bert:qa")     # BERT QA

dataset.save("training.jsonl", format="messages")

{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant..."},
    {"role": "user", "content": "Can I get a refund?"},
    {"role": "assistant", "content": "I'd be happy to help..."}
  ]
}

dataset.save("training.jsonl", format="chatml")

{
  "text": "<|im_start|>system\nYou are a helpful assistant...<|im_end|>\n<|im_start|>user\nCan I get a refund?<|im_end|>\n<|im_start|>assistant\nI'd be happy to help...<|im_end|>"
}

dataset.save("eval.jsonl", format="qa")

{
  "question": "Can I get a refund after 45 days?",
  "answer": "Unfortunately, our refund policy allows returns within 30 days...",
  "expected_outcome": "Politely decline - outside refund window",
  "category": "Refund Policy",
  "scenario_type": "negative"
}

dataset.save("langsmith.jsonl", format="langsmith")

{
  "inputs": {
    "question": "Can I get a refund?"
  },
  "outputs": {
    "answer": "I'd be happy to help with your refund request..."
  },
  "metadata": {
    "category": "Refund Policy",
    "scenario_type": "positive"
  }
}

dataset.save("langfuse.jsonl", format="langfuse")

{
  "input": "Can I get a refund?",
  "expected_output": "I'd be happy to help with your refund request...",
  "metadata": {
    "category": "Refund Policy",
    "scenario_type": "positive"
  }
}

dataset.save("tools.jsonl", format="tool_call")

{
  "messages": [
    {"role": "user", "content": "What's the weather in NYC?"},
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "call_abc123",
        "type": "function",
        "function": {
          "name": "get_weather",
          "arguments": "{\"location\": \"NYC\"}"
        }
      }]
    },
    {
      "role": "tool",
      "content": "72F, sunny",
      "tool_call_id": "call_abc123"
    },
    {"role": "assistant", "content": "The weather in NYC is 72F and sunny."}
  ]
}

dataset.save("bert.jsonl", format="bert")
# or explicitly:
dataset.save("bert.jsonl", format="bert:classification")

{
  "text": "Can I get a refund after 45 days?",
  "label": "negative",
  "category": "Refund Policy"
}

dataset.save("bert_qa.jsonl", format="bert:qa")

{
  "question": "What is the refund window?",
  "context": "Our policy allows returns within 30 days of purchase...",
  "answer_text": "30 days",
  "answer_start": 35
}

from synkro.formatters import BERTFormatter

# Define custom formatter
def my_formatter(trace):
    return {
        "input_text": trace.user_message,
        "target_text": trace.assistant_message,
        "custom_field": trace.scenario.category,
    }

# Register
BERTFormatter.register_task("my_task", my_formatter)

# Use
dataset.save("custom.jsonl", format="bert:my_task")

dataset.save("readable.jsonl", pretty_print=True)

{
  "messages": [
    {
      "role": "user",
      "content": "Can I get a refund?"
    },
    {
      "role": "assistant",
      "content": "I'd be happy to help with your refund request."
    }
  ]
}

from synkro.formatters import (
    MessagesFormatter,
    ChatMLFormatter,
    QAFormatter,
    LangSmithFormatter,
    LangfuseFormatter,
    ToolCallFormatter,
    BERTFormatter,
)

# Format traces
formatter = MessagesFormatter(include_metadata=True)
examples = formatter.format(dataset.traces)

# Save with custom path
formatter.save(dataset.traces, "output.jsonl", pretty_print=True)

# Get JSONL string
jsonl_str = formatter.to_jsonl(dataset.traces)

# Convert to HF Dataset with specific format
hf_dataset = dataset.to_hf_dataset(format="bert:classification")
hf_dataset.push_to_hub("my-org/bert-policy-classifier")

# Messages format for fine-tuning
hf_dataset = dataset.to_hf_dataset(format="messages")
hf_dataset.push_to_hub("my-org/policy-sft-data")