marin-community/

openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16

Synthetic generations from Qwen/Qwen3-4B on the Marin OpenThoughts-4 code SDG prompt set. Each prompt is sampled n=16 times, and for every generated token the dataset stores the chosen-token log probability plus the top-16 log probabilities over the vocabulary, enabling distillation, KL-style fine-tuning, reranking, and uncertainty analysis.

	Generation setup




	

Field Value

Generator model Qwen/Qwen3-4B… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16.