HTMLNLM

Sovereign Browser-Based Compute
LOCAL SANDBOX
NETWORK TOPOLOGY (RWKV-v6)
0.00M
Total Ternary Parameters
CORPUS INGESTION & BPE

Drop .txt corpus here or click
Requires uncompressed UTF-8 text
SYSTEM TELEMETRY
OOMB CHUNK-RECURRENT LOSS
Global Step 0
Current Loss 0.0000
Throughput 0.0 tok/sec
QUANTIZATION DISTRIBUTION (b1.58)
-1 (33%) 0 (34%) +1 (33%)
OPTIMIZER (QUINTIC MUON)
REAL-TIME INFERENCE SAMPLING
Model not allocated. Proceed to ARCHITECTURE tab.
GROUP RELATIVE POLICY OPTIMIZATION

Executes critic-free reinforcement learning. Generates a cohort of responses, evaluates against the target reward condition, and updates parameters via Z-score normalized advantages with an approximate KL divergence constraint.

REINFORCEMENT METRICS
Epochs 0
Cohort Size 4
Avg Reward 0.000
COHORT TRAJECTORIES
Awaiting GRPO initialization...
DECODING SAMPLER
RUNTIME OUTPUT
VIRTUAL MACHINE STATE PERSISTENCE

Master weights residing in Javascript Float32Arrays are serialized into binary formats. Local DB logic utilizes native ArrayBuffer storage for robustness, strictly eliminating Base64 call stack limits. Momentum buffers from Muon are strategically discarded to halve the storage footprint.

LOCAL INDEXED DB

EXTERNAL I/O (JSON)