Datasetpublic

shizhuo2/omega-het-expandA-sft

OMEGA-HET-expandA — matched HET-vs-HOM SFT (equal-size) Matched supervised-fine-tuning data for the OMEGA diversity experiment: for each math prompt, reasoning trajectories are sampled two ways and only prompts solved (math-verified correct) in both conditions are kept (matched HOM∩HET = 3,219 prompts), so HET and HOM are directly comparable. HET (heterogeneous): true token-level continuation across a 3×32B roster (Qwen3-32B + DeepSeek-R1-Distill-Qwen-32B +… See the full description on the dataset page: https://huggingface.co/datasets/shizhuo2/omega-het-expandA-sft.

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes60downloads

Dataset Card

OMEGA-HET-expandA — matched HET-vs-HOM SFT (equal-size)

Matched supervised-fine-tuning data for the OMEGA diversity experiment: for each math prompt, reasoning trajectories are sampled two ways and only prompts solved (math-verified correct) in both conditions are kept (matched HOM∩HET = 3,219 prompts), so HET and HOM are directly comparable.

—HET (heterogeneous): true token-level continuation across a 3×32B roster (Qwen3-32B + DeepSeek-R1-Distill-Qwen-32B + OpenReasoning-Nemotron-32B).
—HOM (homogeneous): a single Qwen3-4B.

Built from the verified pool (shizhuo2/omega-het-expandA-verified; 575,699 accepted: HET 416,727 / HOM 158,972).

Diversity levels (each has matched HET & HOM files)

level	per-prompt selection	HET rows	HOM rows	equal-size
1	best 1 trajectory / prompt	3,219	3,219	yes (count)
K4	4 strategy-diverse / prompt (eligible 2,669)	10,676	10,676	yes (count)
ALL_BALANCED	min(#HOM,#HET) / prompt	124,471	124,471	yes (count)
ALL_RAW (scaling)	all accepted / condition	416,727	158,972	—

Two budget modes per level:

—`_natural` — count-equalized: HET and HOM have the SAME number of rows (the equal-size set).
—`_token_matched` — total-token-equalized: whole trajectories downsampled from the heavier-token side (never trimming math), so total SFT tokens match across HET/HOM (rows may differ).

Schema (per line)

problem_id, condition, generator_type (HET|HOM), diversity_level, axis, setting_key, messages [user prompt, assistant solution], ground_truth, generator_metadata, verification, strategy, token_count

Raw pre-verify pool: shizhuo2/omega-het-expandA-200k-raw • Verified pool: shizhuo2/omega-het-expandA-verified