Zils lite champion relay v1

A full BF16 decision-model body for the Zils subnet 127 d3-lite pool, trained from qhoenix/zils-lite-decision-v9. This is an experimental miner submission. Local evaluation does not establish a paid seat or private-exam performance.

The runtime, tokenizer, answer layer and vision tower are the pinned d3-lite-0.8b-v1 chassis. This repository supplies the body weights; use the pool's pinned chassis to run them. It is not a standalone generic text-generation model.

Trained on one RTX 3090 with merged LoRA rank 64, alpha 128, dropout 0, learning rate 5e-5, AdamW without weight decay, accumulation 16, seed 17, 13,000 training rows / 813 optimizer steps. The official trainer's linear decay and warmup were used. The answer layer stayed fixed. See training.json for pinned revisions, source counts, mixture and limitations.

Nine thousand new rows were reconstructed from the pinned public Zils Relay / tasksource original TRAIN splits; probability targets mix 80% d3-ultra teacher output with 20% original source distribution. Four thousand earlier TRAIN rows were reused for retention, verified synthetic decisions and public outcome prediction. Earlier evaluation rows and matching source groups/prompts were excluded. A fresh group-disjoint broad set was split into candidate selection and final confirmation. No private validator exam data was used. Upstream champion pretraining/fine-tuning exposure to public sources is unknown.

Public proxy evaluation

Scores use the validator's Brier/RPS skill formula, 20% separate temperature calibration and present-cell weight renormalization. These are local proxy scores; they cannot be compared directly with a nightly board score.

evaluation.json preserves all tested development variants, the actual equal-weight champion/candidate soup, five calibration splits, conditional paired bootstrap intervals, final confirmation and structural/reference diagnostics. The submitted native candidate regressed by 0.49 point on the seed-only proxy, with small regressions across all five calibration splits. Its primary development gain was 1.23 points, with a conditional 95% interval of -0.37 to +3.14; that interval includes zero. The primary estimate only narrowly exceeds the policy's 1.2-point fine-tune bar, which the validator will measure on different, private items. The synthetic set shares its generator with part of training and measures transfer within that generator only. The fresh broad confirmation comes from held-out source groups within original TRAIN splits, not an independent official benchmark. No released-model or incentive claim follows from these experiments.

Attribution

Apache-2.0. Original upstream notices and card are preserved in upstream/; see LICENSE and NOTICE. The starting body reports earlier thermopylae-777/zils-lite-body-2 and dasLOL/zils-lite-skills-v5 soup lineage, plus vllm-sr/d3-lite and Qwen/Qwen3.5-0.8B. That upstream training process was not independently reconstructed. Public source text retains its original licences; no source text or held-out examples are redistributed here.