openai · charmquark1984 · Mar 21, 2026
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/README.md b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/README.md
diff --git a/...rack_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/bigram4096_seed1337.txt b/...rack_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/bigram4096_seed1337.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/cudnn2_wd05.txt b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/cudnn2_wd05.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/fact_canon.txt b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/fact_canon.txt
diff --git a/...s/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/gradquant_seed42.txt b/...s/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/gradquant_seed42.txt
diff --git a/...ack_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/int4_L11_mlp3_8xh100.txt b/...ack_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/int4_L11_mlp3_8xh100.txt
diff --git a/...ack_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/memtokens_64_fullcfg.txt b/...ack_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/memtokens_64_fullcfg.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/mtp2_lw03.txt b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/mtp2_lw03.txt
diff --git a/...s/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/pr315_bigram8192.txt b/...s/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/pr315_bigram8192.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/pr315_raw.txt b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/pr315_raw.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/reptile_ttt.txt b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/reptile_ttt.txt
diff --git a/.../track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/stride32_baseline.txt b/.../track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/stride32_baseline.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/ttt_v2_lr01.txt b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/logs/ttt_v2_lr01.txt
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/submission.json b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/submission.json
@@ -0,0 +1,18 @@
+{
+    "author": "charmquark1984",
+    "github_id": "charmquark1984",
+    "name": "Negative Results: 24hrs of 8xH100 Experimentation",
+    "blurb": "Non-record 8xH100 submission: 13 techniques tested that did not work on PR #315 base (causal TTT, MTP, INT4, Canon layers, memory tokens, etc.) plus 4 positive findings (EMA>SWA by 0.003, WD controls artifact size, 786K>524K batch, FA3 Hopper speed). Best verified: 1.1257 BPB.",
+    "date": "2026-03-21T00:00:00Z",
+    "track": "non-record-16mb",
+    "val_loss": 1.90058621,
+    "val_bpb": 1.1257,
+    "pre_quant_val_loss": null,
+    "pre_quant_val_bpb": 1.1427,
+    "step_stop": 6952,
+    "wallclock_seconds": 600,
+    "bytes_total": 15533063,
+    "bytes_model_int8_zlib": null,
+    "bytes_code": null,
+    "gpu": "8xH100_SXM"
+}
diff --git a/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/train_gpt.py b/records/track_non_record_16mb/2026-03-21_Negative_Results_24hr_8xH100/train_gpt.py