Minimal recurrent motif (sb1 rs2 g0.18) – non-record submission

records/track_non_record_16mb/motif_sb1_rs2_g018/README.md.txt

@@ -0,0 +1,35 @@
+Tim Megna
+
+This submission explores a minimal recurrent motif as a structural generator for sequence modeling. The motif was informed by an earlier geometric intuition: that large effective structures can be closed by a small affine walk rather than explicitly constructed.
+
+In particular, the form |3n−2|{-2,5} is treated as an affine generator governing recurrence and reuse. This perspective motivated a compact shared-block architecture capable of producing extended structure through iteration rather than parameter expansion.
+
+Two runs demonstrate that this motif provides stable path-like dynamics and effective closure under simple directional compositions (e.g., U/R/D/L walks), while remaining efficient in parameter usage.
+
+Result
+
+A minimal recurrent motif:
+
+shared_block_size = 1
+recurrence_steps = 2
+recurrence_gate_init = 0.18
+
+Final (roundtrip):
+- val_loss: 4.7062
+- val_bpb: 2.7873
+
+Artifact:
+- compressed: 1.92 MB
+- raw: 8.47 MB
+achieves:
+
+2.787 bpb
+improved compression relative to larger motif variants
+significantly reduced parameter and compute footprint
+
+
+This suggests that effective structure can be generated by a compact shared operator, rather than requiring explicit depth or width.
+
+Notes on Evaluation
+
+Logging and validation timing on the local system exhibit dependence on evaluation chunking (k) and print cadence. This affects perceived runtime during validation (e.g., heartbeat intervals), but does not impact correctness of reported loss or bpb.

records/track_non_record_16mb/motif_sb1_rs2_g018/logs/motif_sb1_rs2_g018_eval_final.txt

@@ -0,0 +1,19 @@
+step:40/80 val_loss:5.3019 val_bpb:3.1401 train_time:5590ms step_avg:139.74ms
+step:45/80 train_loss:5.1820 train_time:6264ms step_avg:139.20ms
+step:50/80 train_loss:5.0359 train_time:6952ms step_avg:139.03ms
+step:55/80 train_loss:5.0466 train_time:7633ms step_avg:138.77ms
+step:60/80 train_loss:4.9535 train_time:8309ms step_avg:138.49ms
+step:60/80 val_loss:4.9555 val_bpb:2.9349 train_time:8310ms step_avg:138.50ms
+step:65/80 train_loss:4.9348 train_time:9111ms step_avg:140.17ms
+step:70/80 train_loss:4.7258 train_time:9923ms step_avg:141.76ms
+step:75/80 train_loss:4.5686 train_time:10613ms step_avg:141.50ms
+step:80/80 train_loss:4.6509 train_time:11352ms step_avg:141.89ms
+step:80/80 val_loss:4.7055 val_bpb:2.7868 train_time:11352ms step_avg:141.90ms
+peak memory allocated: 382 MiB reserved: 408 MiB
+Serialized model: 8413875 bytes
+Code size: 60155 bytes
+Total submission size: 8474030 bytes
+Serialized model int8+zlib: 1862844 bytes (payload:2910240 raw_torch:2916189 payload_ratio:2.89x)
+Total submission size int8+zlib: 1922999 bytes
+final_int8_zlib_roundtrip val_loss:4.7062 val_bpb:2.7873 eval_time:364934ms
+final_int8_zlib_roundtrip_exact val_loss:4.70623326 val_bpb:2.78729643