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 | Stage | Count | | --------------------------------------------- | -------: | | Source prompts from Nemotron FC + Conv | 106,588 | | Successfully sampled trajectories, n=4 | 105,350 | | Total candidate completions | 421,400 | | Auto-scored function-call ground-truth rows | 69,955 | | Rows with at least one chosen candidate | 31,291 | | Rows with a preference pair | 18,822 | | Final SFT rows, full split | 35,850 | | Final SFT rows, no-multistep split | 14,545 | Source Datasets | Tag | Hugging Face Dataset | Source Rows | Rows Kept | | ---- | ---------------------------------------------------------------------- | ----------: | --------: | | FC | `nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1` | 9,620 | 8,390 | | Conv | `nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1` | 96,968 | 96,960 | Sampling Protocol | Field | Value | | ------------------ | ---------------------------------- | | Model | `Qwen/Qwen3.6-35B-A3B-FP8` | | Decoding | `temperature=0.7`, `n=4` | | `top_p` | vLLM default | | `max_tokens` | 12,288 | | `enable_thinking` | `true` | 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. **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 row