ya9919/sharp-421m-mlx
0.4B • Updated • 1
S²-MAS Models · sharp
sharp decision models for S²-MAS: 421M, 322M, 4B and 9B; training variants and MLX runtimes.
Note MLX FP16 runtime of sharp-421m, used in the reported 12M and monthly applications.
ya9919/sharp-322m
0.3B • UpdatedNote 322M bilingual encoder; initial-supervision checkpoint with English and Chinese support.
ya9919/sharp-322m-mlx
0.3B • UpdatedNote MLX runtime of sharp-322m. The paper's A100 single-request timing reaches 22.72× Jev API input-token throughput.
ya9919/sharp-421m
0.4B • UpdatedNote 421M encoder; full six-round SFT + RLCD checkpoint, 2,925,263 training examples.
ya9919/sharp-421m-3rounds
0.4B • UpdatedNote 421M encoder; three-round native checkpoint documentation. No usable weights currently hosted.
ya9919/sharp-421m-3rounds-mlx
0.4B • UpdatedNote Three-round MLX format documentation. No usable weights currently hosted.
ya9919/sharp-4b-head-50k
UpdatedNote 4B model; frozen backbone and trained decision head, 50k examples.
ya9919/sharp-4b-head-200k
UpdatedNote 4B model; frozen backbone and trained decision head, 200k examples.
ya9919/sharp-4b-lora-50k
UpdatedNote 4B model; LoRA plus decision head, 50k examples.
ya9919/sharp-4b
UpdatedNote 4B model; selected LoRA plus decision-head checkpoint, 200k examples.
ya9919/sharp-9b
UpdatedNote 9B model; LoRA plus decision head, 200k examples.
ya9919/sharp-421m-sft-50k
0.4B • Updated • 50Note 421M encoder; SFT on 50k examples. Matched trained version.
ya9919/sharp-421m-sft-200k
0.4B • Updated • 46Note 421M encoder; SFT on 200k examples. Matched trained version.
ya9919/sharp-421m-rlcd-50k
0.4B • Updated • 49Note 421M encoder; RLCD continuation of the matched 50k SFT checkpoint.
ya9919/sharp-421m-rlcd-200k
0.4B • Updated • 53Note 421M encoder; RLCD continuation of the matched 200k SFT checkpoint.
ya9919/sharp-421m-ce-200k
0.4B • Updated • 53Note 421M encoder; additional CE continuation of the matched 200k SFT checkpoint.
ya9919/sharp-421m-compat
0.4B • UpdatedNote Compatibility entry for sharp-421m; does not add an independent trained version.