Skip to content

Commit 388ac59

Browse files
PiyushDattaPiyush Datta
authored andcommitted
Record: SP8192 + Depth Recurrence + Polar Express NS + Phased LoRA TTT
11-layer GPT with SP8192, MLP 4x, depth recurrence (layers 3-5 looped), parallel residuals, Polar Express Newton-Schulz optimizer, SDClip GPTQ (int6 + int8 embed), brotli compression, SWA, and phased LoRA TTT. 3-seed quantized_sliding_window val_bpb mean: 1.09085 PR: openai#2106
1 parent 6097c88 commit 388ac59

1 file changed

Lines changed: 9 additions & 9 deletions

File tree

  • records/track_10min_16mb/2026-04-30_PiyushDatta_SP8192_DepthRecur_PolarNS_LoRATTT

records/track_10min_16mb/2026-04-30_PiyushDatta_SP8192_DepthRecur_PolarNS_LoRATTT/logs_summary.md

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -6,21 +6,21 @@ validation metric is `quantized_sliding_window val_bpb`.
66

77
## Final Metrics
88

9-
| Seed | Train stop step | Last scheduled val step | Final completed val_bpb | Artifact size |
10-
| --- | ---: | ---: | ---: | ---: |
11-
| 42 | 8597 | 8597 | 1.08934733 | 15,999,684 bytes |
12-
| 314 | 8631 | 8631 | 1.09035192 | 15,997,730 bytes |
13-
| 999 | 8620 | 8620 | 1.09285937 | 15,998,747 bytes |
9+
| Seed | Train stop step | Last scheduled val step | Final completed val_bpb | Artifact size |
10+
| ---- | --------------: | ----------------------: | ----------------------: | ---------------: |
11+
| 42 | 8597 | 8597 | 1.08934733 | 15,999,684 bytes |
12+
| 314 | 8631 | 8631 | 1.09035192 | 15,997,730 bytes |
13+
| 999 | 8620 | 8620 | 1.09285937 | 15,998,747 bytes |
1414

1515
## Mean
1616

17-
- `quantized_sliding_window val_bpb` mean: `1.09085287`
17+
- `quantized_sliding_window val_bpb` mean: `1.09085287`
1818

1919
## Source Logs
2020

21-
- `logs/seed_42.log`
22-
- `logs/seed_314.log`
23-
- `logs/seed_999.log`
21+
- `logs/seed_42.log`
22+
- `logs/seed_314.log`
23+
- `logs/seed_999.log`
2424

2525
---
2626

0 commit comments

Comments
 (0)