GLM-5.1-Reasoning-1M-filtered: Refusal-Filtered Instruction-Tuning Data

Jongsim created a filtered version of the GLM-5.1-Reasoning-1M dataset by removing assistant responses containing explicit refusal patterns. The dataset, last updated on April 21, 2026, includes a main subset of 19 GB and a mathematics-focused subset of 4.1 GB. It is a text-based collection intended for training large language models.

Use Cases

• Training language models on refusal-free instruction-response pairs based on the filtered assistant response field.
• Fine-tuning models for mathematical reasoning tasks based on the separate mathematics reasoning subset.
• Benchmarking model performance on instruction-following tasks using a dataset with minimal refusal artifacts.
• Studying the impact of refusal patterns on model training by comparing with the original source dataset.

Strengths

• Filtering applied to remove explicit refusal patterns (e.g., 'I cannot', 'as an AI') from assistant responses.
• Contains a dedicated 4.1 GB subset focused on mathematics reasoning.
• Drop ratio from filtering was approximately 0.00%, indicating the source data had very few refusals.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for specific training needs.
• Description metadata is limited; actual data quality and content diversity require manual inspection after download.

Provenance

Source

Filtered subset of zai-org/GLM-5.1-Reasoning-1M from Hugging Face.

Collection Method

Conservative regex/string match filtering applied to the assistant response field.

Freshness

Last updated 2026-04-21 14:25:58; freshness should be verified.