FunctionGemma 270M + TinyCeNN Memory Fusion

Base revision: 303296b8f3262f08ecdfc6008e94374bbd697d02
Accepted layers: [5, 11, 17]
Status: complete

Only original full-attention anchors are replaced; Gemma3 sliding-window layers remain unchanged.

Layer	Accepted	NMSE	Cosine	Incremental ΔNLL
5	False	0.39167	0.77734	-1.94749
5	False	0.24159	0.86916	-1.92459
5	True	0.15754	0.92262	-2.06301
11	True	0.16000	0.92147	-0.23931
17	False	0.15327	0.92281	+0.07050
17	False	0.11939	0.94842	+0.02532
17	True	0.13087	0.94070	+0.00713

Acceptance gates: NMSE ≤ 0.20, cosine ≥ 0.90, incremental ΔNLL ≤ +0.015, cumulative ΔNLL ≤ +0.05. training/sequential_in_progress.pt is resumable research state and may contain an unaccepted layer.

Source: https://github.com/vtavakkoli/TinyCeNN-LM

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for vtava/functiongemma-270m-it-simple-tool-calling-memory-fusion

Base model

google/functiongemma-270m-it

Finetuned

vtava/functiongemma-270m-it-simple-tool-calling

Finetuned

(1)

this model