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

[ofirkris]

Results

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

• 8xH100 SXM, 600s, 92ms/step
• PyTorch 2.7.0 + FlashAttention 2.8.3
• No TTT, no test-time adaptation

Techniques

10L 512d, 3x MLP, XSA last 4 layers, EMA 0.997, Partial RoPE 16/64, LN Scale,
SmearGate, BigramHash(10240), Int5 MLP / Int6 attn, FP16 embeds, 3.2% pruning, zstd-22,
sliding window eval stride=64

[mohosy]

partial rope is interesting havent seen that in many submissions yet. how many dims did you find works best? 16 seems low but if it works it works

[ofirkris]

partial rope is interesting havent seen that in many submissions yet. how many dims did you find works best? 16 seems low but if it works it works

I tested 16 out of 64 dims (25%) based on ablations from other competitive runs on this challenge.

The intuition is that most heads don't need full positional information - leaving 48 dims position-free lets them learn content-based attention patterns.
I didn't sweep extensively (tried 0, 16, 32, 64) and 16 gave the best result on my setup. Would be interesting to see if anyone finds a different sweet spot at larger layer counts.