Skip to content

10L XSA + EMA + Partial RoPE + LN Scale (val_bpb: 1.1365) #458

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ofirkris wants to merge 5 commits into
base: main
Choose a base branch
from
Open

10L XSA + EMA + Partial RoPE + LN Scale (val_bpb: 1.1365) #458

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e96bad3
submission: XSA + EMA + Partial RoPE + LN Scale + TTT — val_bpb 1.1365
ofirkris Mar 22, 2026
05a4243
fix: remove PR number references from README
ofirkris Mar 22, 2026
74eee56
clean up README
ofirkris Mar 22, 2026
3165236
fix: remove invalid TTT, fix eval model, fix labels
ofirkris Mar 22, 2026
663a135
submission: 10L XSA+EMA+PartialRoPE+LNScale val_bpb=1.1365 (2 seeds)
ofirkris Mar 22, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 27 additions & 0 deletions records/track_10min_16mb/2026-03-22_XSA_EMA_PartialRoPE_LNScale_TTT/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
# 10L XSA + EMA + Partial RoPE + LN Scale

**val_bpb: 1.1365** (mean of seeds 42, 1337)

## Results
| Seed | val_bpb | Artifact | Steps |
|------|---------|----------|-------|
| 42 | 1.1365 | 15,759,319 | 6491 |
| 1337 | 1.1366 | 15,820,386 | 6520 |

## Architecture
- 10 layers, 512d, 8/4 GQA, 3x MLP (1536), ReLU^2
- XSA last 4 layers, EMA 0.997, Partial RoPE 16/64, LN Scale
- SmearGate + BigramHash(10240, 128), U-Net skips, softcap=30

## Training
- BF16, Muon (WD=0.04, momentum=0.99) + AdamW
- 786K batch tokens, seq 2048, warmdown 3000

## Quantization
- Int5 MLP / Int6 attention, FP16 embeds, 3.2% pruning, zstd-22

## Eval
- Sliding window stride=64

## Environment
- PyTorch 2.7.0+cu128, FlashAttention 2.8.3, 8xH100 SXM
12 changes: 12 additions & 0 deletions records/track_10min_16mb/2026-03-22_XSA_EMA_PartialRoPE_LNScale_TTT/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
{
"name": "10L XSA + EMA + Partial RoPE + LN Scale",
"author": "ofirkris",
"github_id": "ofirkris",
"date": "2026-03-22",
"val_bpb": 1.13647704,
"val_loss": 1.91888879,
"artifact_bytes": 15759319,
"training_time_seconds": 600,
"gpu_config": "8xH100_SXM",
"seeds": {"42": 1.13647704, "1337": 1.13662891}
}
Loading