[Non-record] 1xH100 LeakyReLU2 + Legal TTT screening negative result

records/track_non_record_16mb/2026-03-30_1xH100_LeakyReLU2_LegalTTT_screen/README.md

@@ -0,0 +1,26 @@
+# 1xH100 LeakyReLU2 + Legal TTT screening
+
+This is a non-record submission intended as a low-cost 1xH100 screening run of the 2026-03-23_LeakyReLU_LegalTTT_ParallelMuon record family.
+
+## Why submit this
+I used this run to test whether this newer frontier branch still looked promising in a cheaper 1xH100 setting before requesting more compute for multi-GPU reproduction.
+The result was an interesting negative: the run looked reasonable during training, but performance collapsed after the final compressed artifact and final evaluation path.
+
+## Setup
+- Hardware: 1xH100
+- Wallclock cap: 180 seconds
+- Seed: 42
+- Dataset: FineWeb sp1024, train-shards 1
+- Script: copied from the original record folder with no algorithmic changes
+
+## Result
+- Mid-run validation at wallclock stop:
+  - val_bpb: 2.1050
+- Final compressed artifact result:
+  - final_int6_roundtrip_exact val_bpb: 4.62390226
+- Final submission size:
+  - 5019370 bytes
+
+## Takeaway
+This branch showed signs of life during training but did not survive the full final artifact and evaluation path in this low-budget 1xH100 regime.
+I am submitting it as an interesting negative result and screening datapoint, not as a leaderboard attempt.

records/track_non_record_16mb/2026-03-30_1xH100_LeakyReLU2_LegalTTT_screen/submission.json

@@ -0,0 +1,9 @@
+{
+  "submission_name": "1xH100 LeakyReLU2 + Legal TTT screening",
+  "github_username": "schoudhary90210",
+  "author": "Siddhant_Choudhary",
+  "val_bpb": 4.62390226,
+  "track": "track_non_record_16mb",
+  "hardware": "1xH100",
+  "notes": "Low-cost 1xH100 screening run. Interesting negative result: mid-run validation improved, but final compressed artifact evaluation collapsed."
+}