Skip to content

SP8192 + PolarExpressNS + MIN_LR + LQER Asym Rank-4 | val_bpb=1.07302 (3-seed mean) #1977

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
sahiee-dev wants to merge 1 commit into
base: main
Choose a base branch
from
Open

SP8192 + PolarExpressNS + MIN_LR + LQER Asym Rank-4 | val_bpb=1.07302 (3-seed mean) #1977

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
28 changes: 28 additions & 0 deletions ...ds/track_10min_16mb/2026-04-30_SP8192_PolarExpressNS_MinLR_LQERAsymR4/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
# SP8192 PolarExpressNS MinLR LQERAsymR4 Baseline

Baseline run establishing the current score for `SP8192` with `PolarExpressNS`, `MIN_LR`, and `LQERAsymR4`.

This baseline implements a 3-layer depth recurrence mechanism, symmetric row/col normalization in the Muon optimizer, a causal token-only n-gram tilt, parallel residuals in layers 7-10, and split learning rates for early/late parameter layers, alongside targeted fixes for proper MLP normalization independence and PyTorch `no_grad()` TTT evaluation. Soft-Round QAT was stripped to provide a clean baseline for these architectural techniques.

## Validation Performance
The submission yields a 3-seed mean `val_bpb` of **1.07302**.

| Seed | `val_bpb` | Artifact Size (bytes) |
|------|-----------|-----------------------|
| 1337 | 1.07294458| 15,955,411 |
| 42 | 1.07270429| 15,953,179 |
| 2025 | 1.07342419| 15,951,874 |
| **Mean** | **1.07302** | **15,953,488** |

## Compliance & Evaluation
- `train_under_600s`: True
- `artifact_under_16mb`: True
- `eval_under_600s`: True
- `no_pre_quant_ttt_on_val`: True
- `score_first_ttt`: True
- `three_seeds`: True

## Contents
- `train_gpt.py`: Main training script with all modifications.
- `submission.json`: Official track evaluation parameters and scores.
- `train_seed*.log`: Detailed logs for all three seed evaluations.
41 changes: 41 additions & 0 deletions records/track_10min_16mb/2026-04-30_SP8192_PolarExpressNS_MinLR_LQERAsymR4/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,41 @@
{
"author": "sahiee-dev",
"github_id": "sahiee-dev",
"name": "SP8192 PolarExpressNS MinLR LQERAsymR4 Baseline",
"date": "2026-04-30",
"track": "10min_16mb",
"val_bpb": 1.07302,
"val_bpb_std": 0.00030,
"seeds": [1337, 42, 2025],
"seed_results": {
"1337": {"val_bpb": 1.07294458, "artifact_bytes": 15955411},
"42": {"val_bpb": 1.07270429, "artifact_bytes": 15953179},
"2025": {"val_bpb": 1.07342419, "artifact_bytes": 15951874}
},
"hardware": "8xH100 80GB SXM",
"pytorch_version": "2.11.0+cu128",
"technique_summary": "PR #1874 (AjAnubolu) base: SP8192 + SmearGate + AttnOutGate w24 + LoRA-TTT + Phased TTT + Polar Express NS + MIN_LR=0.10 + LQER Asymmetric Rank-4",
"compliance": {
"train_under_600s": true,
"artifact_under_16mb": true,
"eval_under_600s": true,
"no_slot": true,
"no_pre_quant_ttt_on_val": true,
"no_etlb": true,
"no_ngram_cache": true,
"score_first_ttt": true,
"three_seeds": true,
"no_text_normalization": true,
"bpb_on_original_bytes": true
},
"attribution": {
"pr1874_base": "@AjAnubolu (PR #1874)",
"sp8192_base": "@bigbag (PR #1493)",
"smeargate_attn_out_gate": "@MarioPaerle (PR #1667)",
"lora_ttt_improvements": "@renqianluo (PR #1767)",
"phased_ttt": "@jorge-asenjo (PR #1700)",
"polar_express_ns": "PR #1344",
"min_lr_warmdown_floor": "PR #1787",
"lqer_asymmetric_rank4": "PR #1797"
}
}
Loading