ariacompute/afm-de

AFM-D Encoder (afm_de)

On-device System 1 decisions over a typed answer space.
Site: ariacompute.com · Org: ariacompute · Hub: ariacompute/afm-de

AFM-D Encoder is the non-autoregressive track of AFM-D: a Laya-style ModernBERT-large DecisionModel with a MASK option head, RLCD-fine-tuned from convaiinnovations/laya. It scores a text/JSON state against user-supplied options and returns a distribution — not open-ended chat, not TypeSafe Jev.

Primitive	Answer space	Result
Choice	1–255 options	choice, probabilities, confidence
Score	2–10 ordered levels	expected score, distribution, confidence
Noul	false / true	noul = P(true)

Companion Decoder (PEFT LoRA on MiniCPM5-2B): ariacompute/afm-dd.

Model details

Subject id	afm_de
Init / base	convaiinnovations/laya (ModernBERT-large DecisionModel)
Encoder backbone	answerdotai/ModernBERT-large
Context	max_len=1024; shared head_max_len=512 for question + option text
Option text	OPTION_DESC_MAX=96 tokens; long states keep the tail (truncate_left)
High-cardinality Choice	embedding shortlist → one forward pass
Decision temperature	fixed 1.0
Confidence	default normalized entropy; product calib uses noul=max / choice=max / score=max; per-bucket confidence temperatures fitted by ECE grid search
Training	RLCD on option distribution (log + spherical + RPS for Score); LABEL_SMOOTHING=0.02 on hard one-hots
Export	Laya-layout safetensors

Checkpoint layout

model.safetensors
rl_agent_config.json
tokenizer/
encoder/

Intended use

• Local / on-device typed decision scoring (Choice / Score / Noul) over a supplied state.
• Offline eval and JevBench comparison via the AFM-D harness.

How to use

TBD

Eval

Decision T=1.0; calib auto picked noul=max / choice=max / score=max.

	AFM-D Encoder	Base Laya
Agreement	71.1%	55.7%
ECE	0.033	0.141
Brier	0.386	0.605

Task / bucket	n	Agree	ECE
choice (all)	2036	73.8%	0.049
choice:3-5	1360	71.6%	0.059
score:3-5	1019	52.6%	0.095
noul:2	1255	81.8%	0.052

Confidence temperatures (approx.): choice:2 ≈1.9, noul:2 ≈1.75, choice:3-5 ≈1.7, choice:6-10/11+ ≈1.1, score:3-5 =1.0. High-conf errors (conf≥0.7 among wrongs) ≈25%.

JevBench

#	System	Score	Intel.	Calib.	Speed	Acc.	Hard
1	SemIf	83.9	74.9	87.1	90.6	81.0%	61.3%
2	Bespoke Nimble-9B	79.2	72.5	76.2	89.8	79.7%	61.3%
3	NeoHorse-Jev-4B	78.9	63.1	84.6	91.9	72.3%	45.0%
4	Kev-4B	78.3	67.2	76.7	92.9	75.8%	54.1%
5	AFM-D Encoder	45.9	41.0	82.7	94.1	59.3%	38.7%
6	AgentJev-0.6B	41.8	40.0	79.8	88.2	58.0%	36.0%
7	Laya	30.9	36.4	57.7	93.9	53.2%	27.9%

AFM-D Encoder tiers: easy 100%, standard 63.9%, hard 38.7%.

Limitations

• ModernBERT encoder track: strong calibration / speed relative to base Laya, weaker hard-tier Intelligence than larger causal peers on JevBench public-proxy.
• Score buckets remain the weakest local eval slice; prefer targeted data over blind extra epochs when ECE stays high.
• English-centric product corpus; peer hard-label imports are best-effort and may be skipped if missing.
• Must be loaded through AFM-D / Laya DecisionModel code — not a drop-in AutoModelForCausalLM chat checkpoint.

License

MIT

Citation / links

• Org: huggingface.co/ariacompute
• Upstream Laya: github.com/NandhaKishorM/laya
• JevBench: github.com/fstandhartinger/jevbench