datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
DeepSeek-V4-Flash-0731-REAM-Healing-Mix-2048
DeepSeek-V4 Flash REAM Healing Mix 2048
Private, deterministic healing mixtures for the 104-expert REAM-pruned
DeepSeek-V4-Flash-0731-120B checkpoint. It was prepared after direct A/B
generation tests found post-pruning degradation in factual accuracy, language
control, thinking delimiters, and long-code repetition.
seq1024 base config
Rows: 2,048
Maximum sequence length: 1,024 DeepSeek-V4 tokens
Non-padding tokens: 1,581,071
Supervised assistant tokens: 1,092… See the full description on the dataset page: https://huggingface.co/datasets/Baekpica/DeepSeek-V4-Flash-0731-REAM-Healing-Mix-2048.Solar-Open2-120B-A15B-REAM-148E-Healing-Mix
Solar-Open2-120B-A15B-REAM-148E Healing Mix
Baekpica/Solar-Open2-120B-A15B-REAM-148E-BF16 의 REAM 병합 손상을 복구하기 위한
비공개 결정론적 힐링 믹스입니다. 모든 시퀀스는 Solar Open 2 chat template로
사전 렌더·토크나이즈되어 있고 assistant 턴에만 loss 마스크가 열려 있습니다.
왜 만들었나
계보는 upstage/Solar-Open2-250B → REAP(184E) → REAM(148E) 입니다.
REAM 이후 고정 프롬프트 A/B 관찰에서 다음이 확인됐습니다.
한국어 응답에서 반복 붕괴 (예: 동일 3-gram이 출력의 88.6% 차지)
지시 준수·포맷 이탈, typo
도메인·언어에 따라 편차가 큰 열화 (일부 프롬프트는 정상)
손상은 라우터(184→148 centroid slice)와 병합된 expert 가중치에… See the full description on the dataset page: https://huggingface.co/datasets/Baekpica/Solar-Open2-120B-A15B-REAM-148E-Healing-Mix.laurashin_SBF_Trial_Day_15_Prosecutors_Hammer_Bankman-Frieds_Contradictions_With_Reams_of_Eviden