zils-lite-body-16: a six-seed mean LoRA change (attention and Gated-DeltaNet only, factor 1.25) on the current Zils LITE soup

A decision body for the Zils d3-lite soup pool (Bittensor subnet 127), Qwen/Qwen3.5-0.8B family.

Start

The current d3-lite soup: the uniform average of its two members, Charlie320/zils-lite-v9a @93b064048a7aeb7cf0d64236136e12c903e67f40 and qhoenix/zils-lite-decision-v9 @083d5b9aa3d2db48155cd6e7f92ac6f9aa057c65 (both Apache-2.0), rebuilt by us with the official Zils soup code (soup.build); the rebuilt soup hash equals the published one (4d23ec5a54c2ab6d0c16fc493819ef113c789ab8edd1ef8756d41b645d2762a3). Zils body digest of the start: 8bebfcbebee0c6fef293680fc44d0f05c2f05a8c014f37eb055602c157cb58dd. Zils's mining guide allows this ("You may fine-tune the current soup").

Our change

• Six LoRA fine-tunes of that start (seeds 5002, 5012, 5022, 5032, 5042, 5052), one per seed: rank 32 / alpha 64 / no dropout on the attention and Gated-DeltaNet projections of all language layers (MLP projections not trained), merged after training. Learning rate 5e-5, weight decay 1.0, 1 epoch, batch 8, 403 optimiser steps per run, label smoothing 0.05, a 0.1 KL hold to the start on every row, 290 unlabelled anchor prompts held to the start, label loss weighted up on rows the start found hard.
• This body = start + 1.25 x (uniform mean of the six fine-tuned bodies minus the start).
• Training rows: 2,929 rows of our own dataset "lh3500" in the exam's 15 known families (5 categories x choice / yes-no / score): rows written by our team (some use public train-split examples only as topic or style seeds), checked by an independent blind answer and our row guards, and rows made by our own rule-checked programs. No exam split files, no public Zils seed items, no Glaive data, no BIG-bench / MMLU, and no other miner's outputs used as labels.

Files and checks

• model.safetensors sha256 14e867d3c4d4d740dee6b5bf6a53b558b9897dbfaa964bf876633cf4d3e3067e; Zils body digest 5468229fc32f4ceaf3819455991bd54c66273ca5db41fa0204d07fa3eba33fd0.
• Copy screen (official sketch method, three seeds): cosine to the soup 0.99694 (line 0.999); share explained by earlier bodies below 0.30.
• Our local read against the start is small and below our own commit bar; we submit it to learn its real score.

Credit for the starting weights belongs to the authors of the two soup members and the works they build on (see NOTICE). No affiliation with or endorsement by any upstream author is implied.