S²-MAS Models · YN
YN decision models: 322M, 421M, 4B and 9B. GitHub: https://github.com/KataDavidXD/s2-mas
0.4B • Updated • 52 • 1Note MLX FP16 runtime of YN-421m, used in the reported 12M and monthly applications.
ya9919/YN-322m
0.3B • Updated • 58Note 322M bilingual encoder; initial-supervision checkpoint with English and Chinese support.
ya9919/YN-322m-mlx
0.3B • Updated • 44Note MLX runtime of YN-322m. The paper's A100 single-request timing reaches 22.72× Jev API input-token throughput.
ya9919/YN-421m
0.4B • Updated • 56 • 1Note 421M encoder; full six-round SFT + RLCD checkpoint, 2,925,263 training examples.
ya9919/YN-421m-3rounds
0.4B • Updated • 48 • 1Note 421M encoder; three-round native checkpoint documentation. No usable weights currently hosted.
ya9919/YN-421m-3rounds-mlx
0.4B • Updated • 46Note Three-round MLX format documentation. No usable weights currently hosted.
ya9919/YN-4b-head-50k
Updated • 16 • 2Note 4B model; frozen backbone and trained decision head, 50k examples.
ya9919/YN-4b-head-200k
Updated • 15 • 2Note 4B model; frozen backbone and trained decision head, 200k examples.
ya9919/YN-4b-lora-50k
Updated • 16Note 4B model; LoRA plus decision head, 50k examples.
ya9919/YN-4b
Updated • 14Note 4B model; selected LoRA plus decision-head checkpoint, 200k examples.
ya9919/YN-9b
Updated • 14Note 9B model; LoRA plus decision head, 200k examples.
ya9919/YN-421m-sft-50k
0.4B • Updated • 29Note 421M encoder; SFT on 50k examples. Matched trained version.
ya9919/YN-421m-sft-200k
0.4B • Updated • 30Note 421M encoder; SFT on 200k examples. Matched trained version.
ya9919/YN-421m-rlcd-50k
0.4B • Updated • 25Note 421M encoder; RLCD continuation of the matched 50k SFT checkpoint.
ya9919/YN-421m-rlcd-200k
0.4B • Updated • 35Note 421M encoder; RLCD continuation of the matched 200k SFT checkpoint.
ya9919/YN-421m-ce-200k
0.4B • Updated • 28Note 421M encoder; additional CE continuation of the matched 200k SFT checkpoint.
ya9919/YN-421m-compat
0.4B • Updated • 41 • 1Note Compatibility entry for YN-421m; does not add an independent trained version.