rl-rag/dr-tulu-sft-qwen35-tools

DR Tulu SFT Data — Qwen3.5 Tool Format

Converted from rl-rag/dr-tulu-sft-unified to Qwen3.5 multi-turn tool-calling format for LLaMA-Factory SFT training.

Data

• exact_answer_multiturn.json: 2,868 exact-answer examples in LLaMA-Factory sharegpt format
  • Filtered to type=exact_answer only
  • Tool calls in JSON format inside <tool_call> tags (LLaMA-Factory converts to Qwen3.5 XML during tokenization)
  • Tool outputs preserved from original DR Tulu data (Serper search results, page content, Semantic Scholar snippets)
  • System prompt and tool definitions imported from elastic-serving Qwen3Adapter

Tool Name Mapping

DR Tulu Original	Qwen3.5 / elastic-serving
serper_google_webpage_search	web_search
serper_fetch_webpage_content	open_url
semantic_scholar_snippet_search	paper_search

Roles and Masking

Role	Trained on?
system	No (masked)
human	No (masked)
observation	No (masked)
function_call	Yes
gpt	Yes

Citation Format

Final answers use <cite id="ID1,ID2">claim text</cite> for citations and \boxed{answer} for short factual answers.

Usage with LLaMA-Factory

Register in dataset_info.json:

{
  "drtulu_qwen35": {
    "file_name": "exact_answer_multiturn.json",
    "formatting": "sharegpt",
    "columns": {"messages": "conversations", "tools": "tools"},
    "tags": {
      "role_tag": "from", "content_tag": "value",
      "user_tag": "human", "assistant_tag": "gpt",
      "observation_tag": "observation", "function_tag": "function_call",
      "system_tag": "system"
    }
  }
}

Then in your YAML config:

dataset: drtulu_qwen35
template: qwen3  # or qwen3_5_text for Qwen3.5 models