Sign inGet started
Dataset · Hugging Face

TianHongZXY/CHIMERA

CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning CHIMERA is a compact but high-difficulty synthetic reasoning dataset with long Chain-of-Thought (CoT) trajectories and broad STEM coverage, designed for reasoning post-training. All examples are fully LLM-generated and automatically verified without human annotation. Total: 9,225 problems Subjects: 8 Topics: 1,179 🔥 Why CHIMERA? Recent reasoning advances rely heavily on… See the full description on the dataset page: https://huggingface.co/datasets/TianHongZXY/CHIMERA.

Author / organizationTianHongZXY
License (as declared)apache-2.0
Downloads226 (as reported by Hugging Face)
Likes24
Created2 March 2026
Last updated at source17 April 2026
reasoningchain-of-thoughtsynthetic-datallmstempost-training
Embed this entry
<iframe src="https://silytics.si/embed/item/6eff584c-9850-4c34-8ece-240830950195" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Free, no tracking. More about feeds and embeds.

Popularity, last 2 days

226 now (-3 since 2026-10-16). Counts as reported by the source, snapshotted daily by Silytics.

Connected in the knowledge graph

Explore this entry in the graph

Recent changes

  • 16 Oct 2026 Description: CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning CHIMERA is a compact but high-difficulty synthetic reasoning datasetwith long Chain-of-Thought (CoT) trajectories and broad STEM coverage, designed for reasoning post-training. → CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning CHIMERA is a compact but high-difficulty synthetic reasoning dataset with long Chain-of-Thought (CoT) trajectories and broad STEM coverage, designed for reasoning post-training

Changes detected by Silytics between refreshes of the source.

Related papers

More from TianHongZXY

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.