Hugging Face's logo

amewebstudio
/
ananke-sclm-multimodal-earcp-v2

Transformers
English
French
sclm
earcp
multimodal
audio
whisper
memory
online-learning
Model card Files
xet
Community

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

๐Ÿง ๐ŸŽค SCLM-Multimodal with EARCP

Stateful Coherent Language Model with Self-Regulating Ensemble

Key Innovation: EARCP (NOT MoE!)

EARCP (Ensemble Auto-Rรฉgulรฉ par Cohรฉrence et Performance) is fundamentally different from MoE:

Aspect MoE EARCP
Learning Offline Online
Signal Input only Performance + Coherence
Weights Static Adaptive
Diversity Can collapse Floor w_min guarantees diversity
Theory None O(โˆš(T log M)) regret bounds

EARCP Formulas 

P_i,t = ฮฑ_P ยท P_i,t-1 + (1-ฮฑ_P) ยท (-โ„“_i,t)   # Performance EMA
C_i,t = (1/(M-1)) ยท ฮฃ_{j!=i} Agreement(i,j)   # Coherence
s_i,t = ฮฒ ยท P_i,t + (1-ฮฒ) ยท C_i,t            # Combined signal
w_i,t โˆ exp(ฮท_s ยท s_i,t) with floor w_min    # Weight update

Architecture 

Audio Input โ†’ Whisper โ†’ Audio Projector โ”€โ”
                                          โ†“
Text Input โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ†’ SCLM Core
                                          โ”‚
                                    โ”Œโ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”
                                    โ”‚   EARCP   โ”‚
                                    โ”‚           โ”‚
                                    โ”‚ E: Encaps โ”‚
                                    โ”‚ A: Align  โ”‚
                                    โ”‚ R: Revis  โ”‚
                                    โ”‚ C: Coher  โ”‚
                                    โ”‚     +     โ”‚
                                    โ”‚  Online   โ”‚
                                    โ”‚ Learning  โ”‚
                                    โ””โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”˜
                                          โ†“
                                    Text Output โ†’ TTS โ†’ Audio

Parameters

Parameter Value
Latent State 512D
EARCP Components 4 (E,A,R,C)
ฮฑ_P (Performance EMA) 0.9
ฮฑ_C (Coherence EMA) 0.85
ฮฒ (Perf/Coherence balance) 0.7
ฮท_s (Sensitivity) 5.0
w_min (Floor) 0.05
Overhead 2.90%

Author

Mike Amega (Logo) - Ame Web Studio

License

MIT

Downloads last month
39
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for amewebstudio/ananke-sclm-multimodal-earcp-v2

Base model

meta-llama/Llama-3.2-3B-Instruct
Finetuned
(846)
this model

Space using amewebstudio/ananke-sclm-multimodal-earcp-v2 1