Fusion Synth Data S1Kx: Synthetic Completions from 5 LLMs

CohereLabs provides synthetic text completions for the s1K-X training split prompts, generated by five different large language models. The dataset includes outputs from models like GEMMA3-27B-IT, KIMI-K2-INSTRUCT, and QWEN3-235B, sampled at a temperature of 0.3. This collection, last updated in October 2025, is designed for research into model aggregation and training data synthesis.

Use Cases

• Training or fine-tuning language models based on aggregated synthetic completions.
• Comparing response quality and style across different teacher models like gemma3-27b and qwen3.
• Developing algorithms for selecting the best output from multiple model generations.
• Studying the effects of temperature sampling on synthetic data diversity.

Strengths

• Includes completions from five distinct teacher models, providing comparative breadth.
• Outputs are generated with a specified sampling temperature (T=0.3), offering consistency.
• Data is sourced from both cloud (TogetherAI) and locally hosted model instances.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.

Provenance

Source

CohereLabs via Hugging Face.

Collection Method

Synthetic text generation from five specified LLMs for a defined prompt split.

Time Range

null

Freshness

Last updated 2025-10-02 05:39:38; freshness should be verified.

Geography

null