Skip to content

GatedAttn + ValueResid + XSA6 + HedgeMixer + Legal TTT — val_bpb: 1.08965 (3-seed mean) #824

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

Closed
sahiee-dev wants to merge 1 commit into from
Closed

GatedAttn + ValueResid + XSA6 + HedgeMixer + Legal TTT — val_bpb: 1.08965 (3-seed mean) #824

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
31 changes: 31 additions & 0 deletions records/track_10min_16mb/2026-03-25_XSA6_BigramHash4K_HedgeMixer_1.1078/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
# XSA6 + BigramHash4K + GatedAttn + ValueResidual on Hedge Mixer Stack

## Summary

Forked from [@agalimova's PR #720](https://github.com/openai/parameter-golf/pull/720) (XSA6 + BigramHash4K + HedgeMixer baseline at 1.1078).
Added two lightweight architectural modifications that significantly improved performance.

## Architectural Changes

1. **Gated Attention** (`attn_gate`): A per-head learned FP32 scalar (init=1.0) multiplied against the attention output, allowing the model to learn head-specific magnitudes.
2. **Value Residual** (`lambda_v`): A per-block learned FP32 scalar (init=0.0) that injects a fraction of the initial embedding `x0` directly into the residual stream.

Both scalars are registered in `CONTROL_TENSOR_NAME_PATTERNS` to remain in FP32 and bypass GPTQ quantization.

## Results

| Seed | val_bpb | val_loss | eval_time |
|------|---------|----------|-----------|
| 42 | 1.08778131 | 1.83666831 | 504s |
| 1337 | 1.09024766 | 1.84083264 | 503s |
| 2025 | 1.09090710 | 1.84194607 | 506s |
| **mean** | **1.08964536** | **1.83981567** | — |

Artifact: 14,917,177 bytes (14.9MB). All seeds evaluated under 600s.

## Evaluation

- `EVAL_STRIDE=64` (matches official baseline default)
- All runs completed in ~503–506s (under the 600s hard limit)
- Hardware: 8×H100 SXM 80GB
- Compression: zstandard (zstd level 22)
Loading