output-format.md

Output Format Reference

BatchGen output follows the OpenAI Batch API format. Each line in the output JSONL file is a BatchResultItem.

Base Structure

{
  "id": "batch_req_<hex>",
  "custom_id": "<from_input_file>",
  "response": {
    "status_code": 200,
    "request_id": "req_<hex>",
    "body": { ... }
  },
  "error": null
}

Chat Completion Response (/v1/chat/completions)

{
  "id": "chatcmpl-<hex>",
  "object": "chat.completion",
  "created": 1234567890,
  "model": "deepseek-ai/DeepSeek-R1",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "The answer is 42.",
        "reasoning_content": null,
        "tool_calls": null
      },
      "logprobs": null,
      "finish_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 128,
    "completion_tokens": 64,
    "total_tokens": 192
  },
  "system_fingerprint": null
}

Text Completion Response (/v1/completions)

{
  "id": "cmpl-<hex>",
  "object": "text_completion",
  "created": 1234567890,
  "model": "deepseek-ai/DeepSeek-R1",
  "choices": [
    {
      "index": 0,
      "text": "The decoded output text...",
      "logprobs": null,
      "finish_reason": null
    }
  ],
  "usage": { ... }
}

Note: --parse-thinking and --parse-tool-call only apply to chat completion responses. Text completion responses always return raw decoded text.

---

Structured Output Fields

By default, the raw decoded text (including any special tokens) is placed in content. The following server flags enable server-side parsing into structured fields.

--parse-thinking

Extracts model reasoning/thinking into a separate reasoning_content field. The content field contains only the visible answer.

Without --parse-thinking:

{
  "content": "<think>Let me reason step by step...</think>The answer is 42.",
  "reasoning_content": null
}

With --parse-thinking:

{
  "content": "The answer is 42.",
  "reasoning_content": "Let me reason step by step..."
}

--parse-tool-call

Extracts tool/function calls into an OpenAI-compatible tool_calls array.

Without --parse-tool-call:

{
  "content": "Let me check.<｜tool▁call▁begin｜>get_weather\n{\"city\": \"London\"}<｜tool▁call▁end｜>",
  "tool_calls": null
}

With --parse-tool-call:

{
  "content": "Let me check.",
  "tool_calls": [
    {
      "id": "call_abc123def456",
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": "{\"city\": \"London\"}"
      }
    }
  ]
}

---

Supported Model Formats

Each model uses different special tokens for thinking and tool calls. BatchGen's tokenizer classes handle the model-specific parsing automatically.

Model	Thinking Format	Tool Call Format
DeepSeek-R1	<think>...</think>	<｜tool▁call▁begin｜>name\n{args}<｜tool▁call▁end｜>
GPT-OSS-120B	<|channel|>analysis<|message|>...	<|call|>{json}
Kimi-K2.5	<think>...</think>	Not yet supported
GLM-5	<think>...</think>	<tool_call>name<arg_key>k</arg_key><arg_value>v</arg_value></tool_call>
MiniMax-M2.5	<think>...</think>	<minimax:tool_call><invoke name="...">...</invoke></minimax:tool_call>

---

Parsing Results in Python

import json

with open("results.jsonl", "r") as f:
    for line in f:
        result = json.loads(line)
        custom_id = result["custom_id"]
        msg = result["response"]["body"]["choices"][0]["message"]

        # Visible content
        print(f"[{custom_id}] {msg['content']}")

        # Reasoning (when --parse-thinking is enabled)
        if msg.get("reasoning_content"):
            print(f"  Reasoning: {msg['reasoning_content'][:100]}...")

        # Tool calls (when --parse-tool-call is enabled)
        if msg.get("tool_calls"):
            for tc in msg["tool_calls"]:
                fn = tc["function"]
                print(f"  Tool call: {fn['name']}({fn['arguments']})")

---

See Also

• Server Flags - Enable --parse-thinking and --parse-tool-call
• Client API - Submitting batches and downloading results