production checkpoint
Generate with Phase 1 Weights
Loading pretrained checkpoint status.
Kyberne Lab / EventMix
Bayesian Predictive Event Transformer
Learn from next-token prediction error inside an isolated session while uncertainty-gated local adapters update across four layers and the pretrained causal backbone remains frozen.
8 causal layers
12 attention heads
15.1M slow parameters
8,192 memory event capacity
loaded model
96x3 / Context 64
- Checkpoint
- step 4,750
- Validation loss
- 7.258231
- Validation accuracy
- 9.96%
- Event rate
- 0.14816
- Optimizer state
- No
- Online updates
- No
GPT-2 BPE Local Attention Spiking Memory SwiGLU Token Readout
model output
Completion
The future of learning is ready for generation.
token distribution
Top Candidates
No token distribution available.
live state
Runtime Signals
- Prompt tokens
- -
- Spike rate
- -
- Attention gate
- -
- State memory
- -
- Plastic connections
- 8,192
- Device
- CPU
benchmark controls
Run EVM-2 v0.25
Ready. Run the EVM-2 v0.25 checkpoint benchmark.
checkpoint and reference
MNIST 32x32 Spiking LeNet Path
Dataset
The verified checkpoint uses MNIST IDX handwritten digits only. Each 28x28 digit is normalized into a 32x32 retina input; no alphabet or symbol dataset is used offline.
Online scope
The lab keeps the MNIST-trained visual hierarchy fixed enough for comparison, while the interactive adapter can absorb user-drawn letters and symbols as online learning samples.
Generalization test
The same sparse visual evidence is projected into an 8x12 ASCII matrix output. New classes are learned incrementally, making the page a direct test of post-training generalization.
pretrained model
evm-2-v0.25-mnist-92p0
- Protocol
- evm-2-mnist-spiking-lenet-v1
- Architecture
- Retina -> S1 -> C1 -> C2 -> Detector -> Part -> Centroid
- Small validation
- 277 / 300, 92.3%
- Large validation
- 920 / 1000, 92.0%
interactive adapter
Online ASCII Test / Train
Draw a digit, letter, or symbol, then test or train it online.
latest result
92.3%
- Correct
- 277 / 300
- Elapsed
- -
- Composable part memory
- 100.0%
- Active assembly rate
- 25.4%
- ASCII matrix readout
- top-2 blend 0.5
- Protocol
- evm-2-mnist-spiking-lenet-v1
network path
Retina -> S1/C1/C2 -> Part Memory -> Centroid
Retina 1024 S1 12 maps C1 3072 C2 768 Detector 64 Part 96 Centroid 320 Output 8x12
confusion matrix
Actual vs Predicted
actual architecture
EVM-2 Runtime Forward Graph
retina / feature maps descriptor prototype banks centroid readout
token corpus
Train Sentence Next-Token Memory
Ready. Train GEMN2 token SNN on sentence continuations.
latest GEMN2 metrics
-
- Top-1
- -
- Top-5
- -
- Candidate recall
- -
- Avg candidates
- -
- Vocabulary
- -
- Trained tokens
- -
evaluation
Test Sentence Prediction
Step 1 Pick a corpus
Step 2 Train token SNN
Step 3 Test next token
Step 4 Predict top-k
No test run yet.
next-token prediction
Input Sentence, Predict Next Token
No prediction yet.
session controls
Reset or Run Token Benchmark
token benchmark
Before / Train / After
- Protocol
- gemn2-token-snn-v0
- Train top-5
- -
- After top-1
- -
- Prediction
- -
actual architecture
GEMN2 Token SNN Runtime Graph
BPE token / embedding spikes stateful SNN traces candidate next-token decoder
bayesian predictive event language model
Learn Locally, Predict with Posterior Confidence
Connecting to the continual Transformer service.
generated text
Autoregressive Output
No generation yet.
slow logits + fast / durable memory
-
runtime state
event-transition-service-v1
- Inference mode
- -
- Returned next token
- -
- Prompt tokens
- -
- Generated tokens
- -
- Learned tokens
- -
- Memory entries
- -
- Durable entries
- -
- Bayesian observations
- -
- Posterior variance
- -
- Fast weight norm
- -
- Session state
- -
- Elapsed
- -
self-supervised deployment checkpoint
Offline Slow Weights. Online Local Bayesian Plasticity.
- Tokenizer
- -
- Gradient parameters
- -
- Model buffers
- -
- Selected checkpoint tokens
- -
- Total training exposure
- -
- Completed steps
- -
- Best step
- -
- Raw validation accuracy
- -
- Validation PPL
- -
- Architecture gate
- -
- Continual learning gate
- -
bayesian-predictive-event-transformer-v5
Language Backbone with Multi-Layer Local Bayesian Learning
Byte BPE2,048 learned symbols
Slow Transformer8 causal layers, 12 heads, RoPE and SwiGLU
Causal Successor Pathgeneric within-context token binding and sequence reuse
Bayesian Local Adaptersrank-16 full-covariance updates at layers 2, 4, 6 and 8
Fast + Durable Memorysession delta updates and validated long-term promotion
Uncertainty-Gated Decodeslow logits + successor evidence + local posterior corrections
- Model dim
- 384
- Transformer layers
- 8
- Attention heads
- 12
- Context
- 384 tokens
- Slow parameters
- 15,096,197
- Plastic layers
- 2 / 4 / 6 / 8
- Bayesian rank
- 16, full covariance
- Memory capacity
- 8,192 events
- Memory key
- 128 dimensions
- Per-session state
- 5.34 MB
Offline next-token pretraining learns language and routing features without concept labels. Online learning freezes the backbone, updates fast weights and four rank-16 Bayesian posteriors from local prediction error, and suppresses corrections in uncertain directions. Stable evidence can be retained in a user-isolated durable store. This is a small research model, not an LLM-scale general assistant.