zake7749/Qwen3.6-35B-A3B-Tool-Calling

Qwen3.6-35B-A3B Tool-Calling Dataset

This repository presents a function and tool-calling preference and supervised fine-tuning dataset constructed from Nemotron-RL agentic prompt corpora.

For each source prompt, the model was sampled four times with thinking mode enabled. Each resulting candidate trajectory was then evaluated against the dataset’s ground-truth action using exact matching on both the function name and the parsed function arguments.

Overview

[image]

Source Datasets

[image]

Sampling Protocol

Difficulty Signal: How Often Does Qwen3.6 Produce the Correct Tool Call?

For the 69,955 auto-scored function-call rows, each prompt was sampled four times. The distribution of chosen candidates per prompt is shown below.

[image]

• —0 / 4 — 55.3%: all four candidates failed to match the ground-truth action. These prompts do not contribute SFT rows, but they may be valuable as negative examples for DPO when paired with the gold action.
• —1–3 / 4 — 26.9%: these rows form the main pool of preference pairs, totaling 18,822 rows.
• —4 / 4 — 17.8%: all four candidates matched the ground truth. One candidate is retained per row as a high-confidence SFT example.

These results indicate that rejection sampling against gold tool-call actions remains challenging for Qwen3.6 on the Nemotron function-calling corpus. The task is therefore substantially more difficult than simple paraphrase generation.

Quality Signal: Chosen Reasoning Is Shorter Than Rejected Reasoning

Across 279,820 auto-scored candidates, chosen trajectories tend to contain shorter reasoning traces than rejected ones.

[image]

Rejected reasoning traces are approximately 76% longer on average than chosen ones.

In addition, finish_reason == "length" appears in 1,164 candidates, representing 0.28% of the scored candidates. These cases occur almost entirely among rejected candidates, suggesting that lower-quality trajectories are more likely to exhaust the token budget.

This aligns with a common observation for Qwen-style thinking models: when the model loses track of the task, its reasoning often becomes longer and less decisive. Training on the chosen subset may therefore naturally bias the student model toward shorter and more reliable tool-calling behavior.

SFT Split Complexity

The cleaned SFT parquet files contain the full conversation prefix and the selected final assistant turn, including reasoning, content, and tool_calls.

[image]

Key characteristics include:

• —The median tool inventory per row is 17, with a maximum of 46. This indicates that most prompts expose the model to a realistic tool ecosystem rather than a small toy setup.
• —The median trajectory length, including system and tool turns, is 7, while the 99th percentile is 28. Multi-turn tool-use sequences are well represented, though they are not the dominant case.

SFT Files

Two supervised fine-tuning parquet files are provided.

The no-multistep split is intended to reduce the risk of leaking prior-turn reasoning into the loss target during training.

Both files share the same schema:

columns: ["tools", "messages"]

Both columns are stored as JSON strings. Only the final assistant turn is used as the supervised fine-tuning target; preceding turns are provided as conversational context.

Selection Rule for Function-Call Rows

Raw Dataset Schema

qwen3.6-35b-a3b_samples.jsonl

The raw rejection-sampling dump contains 105,350 rows. Each row follows the structure below.

{
  "trajectory_id": 3344,
  "dataset": "function-calling-pivot",
  "tools": "[{...}, {...}]",
  "messages": [
    {"role": "system", "content": "..."}
  ],
  "expected_action": {
    "type": "function_call",
    "content": "..."
  },
  "candidates": [
    {
      "index": 0,
      "reasoning": "...",
      "content": "...",
      "tool_calls": [
        {
          "id": "...",
          "type": "function",
          "function": {...}
        }
      ],
      "finish_reason": "tool_calls",
      "evaluation": "chosen"
    }
  ],
  "meta": {
    "num_chosen": 2,
    "num_rejected": 2,
    "num_unscored": 0,
    "has_preference_pair": true
  }
}

For message-type examples, expected_action.content contains the gold response. For function_call examples, it contains the JSON-stringified ground-truth function call.

Cleaned SFT Schema

qwen3.6-35b-a3b_sft_*.parquet

The cleaned SFT files contain two JSON-stringified columns:

columns: ["tools", "messages"]

The messages field follows this structure:

[
  {"role": "system", "content": "..."},
  {"role": "user", "content": "..."},
  {
    "role": "assistant",
    "reasoning": "...",
    "content": "",
    "tool_calls": [...]
  },
  {
    "role": "tool",
    "tool_call_id": "...",
    "content": "..."
  },
  ...
  {
    "role": "assistant",
    ...
  }
]

Only the final assistant turn is used as the loss target. All preceding turns serve as conversational context.

Known Limitations and Caveats

1. 1.Conversational ground truth is not auto-scored. The 35,395 message-type rows produce only unscored candidates. These rows are retained in the raw JSONL file for completeness, but they do not contribute to the SFT splits. Users who wish to incorporate these examples should run a separate LLM-judge evaluation pass.

1. 1.Rows with zero chosen candidates are excluded from SFT. The 0 / 4 rows are dropped from the SFT files. However, they may be especially valuable as negative examples for DPO. Users interested in DPO training should regenerate negatives directly from samples.jsonl rather than relying on the SFT parquet files.

License and Attribution

The source prompts are derived from NVIDIA datasets released under the NVIDIA Open Data License v1.0. Please refer to the original Nemotron-RL Hugging Face repositories for the full license terms.

The Qwen3.6 completions were generated using Qwen/Qwen3.6-35B-A3B-FP8 through local vLLM inference. Downstream usage should follow the terms specified in the Qwen model card.

This dataset, including its layout, scoring procedure, and chosen/rejected annotations, is released under CC-BY-4.0.

When publishing results based on this dataset, please cite both the NVIDIA source datasets and the Qwen3.6-35B-A3B model card.