PR #990

open

ClownCar: Frugendorff compression baseline + canonical DeltaNet integration

val_bpb

0.7614

Architecture

Transformer

Optimizer

—

Artifact Size

9.06MB

Training Techniques

Architecture

depth recurrence

Uses one shared crawler block executed repeatedly across loops, with unique flat encoder/decoder layers around it.

parameters: {"loops":4,"flat_layers":4}

DeltaNet

Canonical DeltaNet integration via CanonicalDeltaNet and chunk_delta_rule.

parameters: {"heads":4}

weight tying

Shares crawler block weights across repeated loop executions.

parameters: null

Quantization

int6

bits: 6

scope: all

Compression

zstd

level: null

Evaluation

sliding window eval

parameters: null

LR Schedule

warmdown

parameters: {"warmdown_iters":2000}

Other

other

Int8 crawler quantization mode used to improve quantization resilience.

parameters: {"env_var":"CRAWLER_QUANT_INT8=1"}

Frugendorff (F-Wing) crawler baseline with a shared recurrent block repeated across loops
Loop-specific instruction perturbations recomputed from the current hidden state each loop
Canonical DeltaNet integration using chunk_delta_rule
Empirical signal analysis separating width effects from weight sharing effects
Int8 crawler quantization mode to mitigate post-processing degradation
Confirmed sub-16MB submission with ~9.06MB artifact size