glm-4.7-Superior-Reasoning-stage1

# glm-4.7-Superior-Reasoning-stage1 ## Dataset Summary `glm-4.7-Superior-Reasoning-stage1` is a Stage1 reasoning distillation dataset built from the **Alibaba Superior-Reasoning** style pipeline, with a stronger teacher model replacement. Compared with the original upstream setup, this release uses **GLM-4.7** as teacher for higher-quality reasoning traces. ## Stage1 Distillation Setup (Low Temperature) - **Training stage:** `stage1` - **Sampling temperature:** **`0.6`** (low-temperature distillation) - `top_p`: `0.95` - Domain focus: `math` - Output records: `1,192` ## Method Note (Based on Alibaba Superior-Reasoning) This card follows the Stage1-style reasoning distillation direction from Alibaba Superior-Reasoning (including the staged reasoning data construction perspective), while replacing the teacher with a stronger model (`glm-4.7`) in this run. ## Data Structure Each line is a JSON object: ```json { "id": "string", "conversation": [ {"from": "human", "value": "..."}, {"from": "gpt", "value": "<think>...</think>\n..."} ], "input": "...", "output": "<think>...</think>\n...", "domain": "math", "meta": { "training_stage": "stage1", "sampling_temperature": 0.6, "teacher_model": "glm-4.7" } } ``` ### Fields - `id`: normalized question hash id - `conversation`: single-turn conversation packaged in chat format - `input`: user question text - `output`: assistant reasoning + answer text - `domain`: task/domain label (mainly `math`) - `meta.training_stage`: stage tag (`stage1`) - `meta.sampling_temperature`: distillation temperature (`0.6`) - `meta.teacher_model`: teacher id (`glm-4.7`) ## Intended Use - Stage1 reasoning SFT - Math CoT behavior tuning - Teacher-model swap ablation studies ## Limitations - This release is distillation data, not ground-truth proof data - Reasoning style inherits teacher preferences - Upstream licensing and usage constraints of source datasets must be respected ## Token and Cost Summary Input tokens: 131,883 Output tokens: 22,608,190 Total tokens: 22,740,073 Cost: $59.98 (Input $0.07 + Output $59.91) ## Project Status Z.ai currently does not provide access to the stronger GLM-5 model, and the distillation pipeline is being migrated. ## References - Alibaba-NLP. *Alibaba-Superior-Reasoning* (dataset card): https://huggingface.co/datasets/Alibaba-NLP/Alibaba-Superior-Reasoning - Li et al. *DASD: Deep Analyze Soon Deliver for Efficient Reinforced Reasoning*: https://arxiv.org/abs/2504.13994 - Wang et al. *SuperiorReasoning: Towards Large-Scale Inference-Time Scaling for Long-CoT Reasoning*: https://arxiv.org/abs/2504.07955