Post
2,523
S²-MAS: 500M household personas • 2,925,263 post-training examples • 12M distinct model-inferred agents / 24M choices.
Main policies: FinQ (4B LoRA + head, 200k) and FinL (421M Laya, six rounds). FinL-mlx is the MLX version.
English FinL-mlx (six-round FP16): 74.5544% accuracy / 38.3746% question-macro F1.
FinL-bilingual-mlx: 9.10 ms latency; 22.72× Jev 1.13 API throughput / 2.42× the same checkpoint in PyTorch.
FinL has the largest domain post-training set among the paper's financial decision models with disclosed scales: 2,925,263 examples.
12M distinct synthetic households form the largest directly model-evaluated population among the compared LLM-inference workloads.
24M choices + 288M monthly ledger states in 3h59m17s on 14 GPUs (8 H100 + 6 A100).
Test: 240,458 records / 119 question versions. Timing: 3 Chinese inputs, 60 calls/arm, batch/concurrency 1, one A100 40GB; API round trip included.
6 routes: Encoder option scoring; LM + learned decision head; Next-token decision SFT; Dynamic schema classification; Multimodal encoder decisions; Provider decision interfaces.
10 development-selected trained variants:
Native: SFT50; SFT200; SFT50→RLCD+50k; SFT200→RLCD+200k; SFT200→CE+200k.
Qwen: 4B Head50; 4B Head200; 4B LoRA50; FinQ (4B LoRA200); 9B LoRA200.
12 benchmarked models/systems; 17 repository entries, including 2 documentation-only three-round entries.
8 financial tasks; 10,000 households over 24 months connect five shocks to purchases, cash, payments and bank offers. Independent synthetic companion: 2M states + 32k unlabeled inputs.
Blog: https://huggingface.co/blog/ya9919/s2-mas
Paper: https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf
Scientific scripts: https://github.com/KataDavidXD/s2-mas
Models: https://huggingface.co/collections/ya9919/s-mas-models-finq-and-finl
Datasets: https://huggingface.co/collections/ya9919/s-mas-dataset-resources
GENESE engine remains private. Component research terms apply; Wikidata CC0.
Main policies: FinQ (4B LoRA + head, 200k) and FinL (421M Laya, six rounds). FinL-mlx is the MLX version.
English FinL-mlx (six-round FP16): 74.5544% accuracy / 38.3746% question-macro F1.
FinL-bilingual-mlx: 9.10 ms latency; 22.72× Jev 1.13 API throughput / 2.42× the same checkpoint in PyTorch.
FinL has the largest domain post-training set among the paper's financial decision models with disclosed scales: 2,925,263 examples.
12M distinct synthetic households form the largest directly model-evaluated population among the compared LLM-inference workloads.
24M choices + 288M monthly ledger states in 3h59m17s on 14 GPUs (8 H100 + 6 A100).
Test: 240,458 records / 119 question versions. Timing: 3 Chinese inputs, 60 calls/arm, batch/concurrency 1, one A100 40GB; API round trip included.
6 routes: Encoder option scoring; LM + learned decision head; Next-token decision SFT; Dynamic schema classification; Multimodal encoder decisions; Provider decision interfaces.
10 development-selected trained variants:
Native: SFT50; SFT200; SFT50→RLCD+50k; SFT200→RLCD+200k; SFT200→CE+200k.
Qwen: 4B Head50; 4B Head200; 4B LoRA50; FinQ (4B LoRA200); 9B LoRA200.
12 benchmarked models/systems; 17 repository entries, including 2 documentation-only three-round entries.
8 financial tasks; 10,000 households over 24 months connect five shocks to purchases, cash, payments and bank offers. Independent synthetic companion: 2M states + 32k unlabeled inputs.
Blog: https://huggingface.co/blog/ya9919/s2-mas
Paper: https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf
Scientific scripts: https://github.com/KataDavidXD/s2-mas
Models: https://huggingface.co/collections/ya9919/s-mas-models-finq-and-finl
Datasets: https://huggingface.co/collections/ya9919/s-mas-dataset-resources
GENESE engine remains private. Component research terms apply; Wikidata CC0.