Within the evolving panorama of AI, the standard of coaching information usually defines the efficiency ceiling of an LLM. Artificial reasoning datasets seize superior drawback‑fixing traces from chopping‑edge fashions — distilling advanced reasoning processes into structured samples. This permits builders to refine their fashions in order that they generate coherent, step‑by‑step explanations and options, bridging the hole between human reasoning and machine effectivity.
1. R1‑Distill‑SFT
- Hyperlink: ServiceNow‑AI/R1‑Distill‑SFT
- Overview:
This intensive assortment options roughly 1.7 million samples distilled from the DeepSeek‑R1‑Distill‑Qwen‑32B mannequin. It aggregates reasoning traces from 9 distinct supply datasets. - Drawback Domains:
Superb for duties in arithmetic, coding challenges, and logical puzzles, it serves as a complete basis for basic‑function reasoning. - Utility:
Use this dataset to nice‑tune massive‑scale LLMs (e.g., these based mostly on Llama or Qwen architectures) aiming to boost general reasoning capabilities throughout a number of domains. - Instance:
