chenghuzi/glimmer-e4b-asd9-gguf

Glimmer E4B ASD-9 GGUF

Glimmer is a Gemma 4 E4B based GGUF model for observable behavior-label prediction. This private preview release is intended for research and screening-support workflows. It is not an ASD diagnostic system and must not be used as the sole basis for diagnosis, treatment decisions, or emergency decisions.

Files

• model-Q4_K_M.gguf: Q4_K_M GGUF language model.
• mmproj-bf16.gguf: BF16 multimodal projection file. Use it together with the main GGUF file.
• metrics.json: Sanitized metric summary only. It does not include predictions, logs, media, cache files, or local paths.

Output Contract

The model is expected to emit a strict 9-bit code for B01 through B09.

^[01]{9}$

B10 is not directly predicted by the model. Applications should derive B10=true only when all B01 through B09 bits are 0.

Labels

Code	Label
B01	Absence or Avoidance of Eye Contact
B02	Aggressive Behavior
B03	Hyper- or Hyporeactivity to Sensory Input
B04	Non-Responsiveness to Verbal Interaction
B05	Non-Typical Language
B06	Object Lining-Up
B07	Self-Hitting or Self-Injurious Behavior
B08	Self-Spinning or Spinning Objects
B09	Upper Limb Stereotypies
B10	Background, application-derived

Evaluation Summary

Evaluation used the unchanged test split with 182 samples. The published metrics use the native-equivalent llama.cpp server baseline with pinned media marker <__media__> and prompt cache disabled for iOS/native parity. The run used Chinese prompts, audio enabled, up to 32 frames, image width 512, max audio 30.0 seconds, deterministic decoding, and grammar-constrained 9-bit output.

Metric	Value
Parse rate	1.0000
Exact match	0.3626
Hamming accuracy	0.8692
Micro precision	0.6059
Micro recall	0.4965
Micro F1	0.5458
Macro precision	0.6461
Macro recall	0.5041
Macro F1	0.5473

Per-Label Metrics

Code	Label	Precision	Recall	F1	Support
B01	Absence or Avoidance of Eye Contact	0.7895	0.3750	0.5085	40
B02	Aggressive Behavior	0.9375	0.4688	0.6250	32
B03	Hyper- or Hyporeactivity to Sensory Input	0.5714	0.3200	0.4103	25
B04	Non-Responsiveness to Verbal Interaction	0.4375	0.2593	0.3256	27
B05	Non-Typical Language	0.4286	0.3750	0.4000	24
B06	Object Lining-Up	1.0000	0.7778	0.8750	9
B07	Self-Hitting or Self-Injurious Behavior	0.3571	0.5263	0.4255	19
B08	Self-Spinning or Spinning Objects	0.7222	0.5200	0.6047	25
B09	Upper Limb Stereotypies	0.6000	0.5660	0.5825	53
B10	Background	0.6170	0.8529	0.7160	34

llama.cpp Usage Notes

Use both GGUF files. The exact command depends on the local llama.cpp build and multimodal support. Keep decoding deterministic and constrain output to the 9-bit grammar when possible.

Example grammar:

root ::= bit bit bit bit bit bit bit bit bit
bit ::= "0" | "1"

License

The GGUF model files are distributed under Apache-2.0. See LICENSE and NOTICE.