zils-lite-ft1

A Zils d3-lite soup-pool body (Bittensor subnet 127): the pinned d3-lite chassis layout (Qwen3.5-0.8B language and vision tensors, BF16), fine-tuned from the d3-lite soup of round d3-lite-2026-11-06, the uniform average of 1lock/zils-lite-astrid-g1 at revision ac979f3d and bmars8218/zils-lite-body-2 at revision 60b596b3 (both Apache-2.0). It is evaluated on the pool's pinned chassis, which supplies the answer layer, prompt and runtime.

Method

• Start: the two soup members averaged and placed on the pinned chassis with the Zils soup builder (zils.soup.build), the way validators build the round's soup.
• Training: one epoch of LoRA (rank 32, alpha 64, learning rate 2e-4) on the language projections that the Zils d3 runtime adapts, with its fixed answer readout (zils.teacher train --start), merged into the weights in BF16. No other tensor was trained. One RTX 3090 (24 GB), about 15 minutes.
• Targets: the verified answer of each item.
• Selection: learning rates 5e-5, 1e-4, 2e-4 and 4e-4 were trained the same way; this one scored best on the local evaluation below.

Training data

4,992 exam-format decision items written by Claude Opus 5.5 with the Zils exam authoring recipe (scripts/author_exam.py: one public row as a style example, an independent blind answer, kept only on agreement), using public training splits only:

• 2,974 items for the 15 current known families: Social IQa, PIQA, SVAMP and GSM8K, GLUE MNLI, RTE, QNLI and STS-B, Crisscrossed Captions (validation), 20 Newsgroups, Glaive function calling v2, dair-ai/emotion, News Headlines Dataset for Sarcasm Detection, and TweetEval sentiment;
• 2,018 items across 99 BIG-bench training tasks (tasksource/bigbench) and Glaive function calling.

No validator exam items, no public seed items and no customer data were used. This repository holds model weights only, no training text; check each source's terms before reusing it.

Local evaluation (not the validators' exam)

1,191 held-out items, scored with the validators' calibrated Brier skill (soup_grade.evaluate, mean of five calibration splits): 892 written the same way from rows never used in training, including 15 BIG-bench tasks kept out of training, plus the 299 current-family items of zils-ai/base-exam-seed-v1 to -v4.

	Start (the soup)	This body
All 1,191 items	43.17	50.82
Public seeds as the known block, held-out tasks as the new block	44.31	51.07

On the same 1,191 items astrid-g1 alone scores 38.84 and body-2 alone 38.31. Validator scores are on a private exam and will differ.

Licence

Apache-2.0, as the start bodies (1lock/zils-lite-astrid-g1; bmars8218/zils-lite-body-2, itself built on thermopylae-777/zils-lite-body-2, dasLOL/zils-lite-skills-v5 and vllm-sr/d3-lite) and Qwen/Qwen3.5-0.8B. Credit to their authors; please cite BIG-bench (Srivastava et al., 2022) and tasksource (Sileo, 2024) for the training sources.