zils-lite-body (v5b)

A fine-tune of qhoenix/zils-lite-decision-v9 (revision 083d5b9a, the current Zils d3-lite soup; itself a fine-tune of the soup of thermopylae-777/zils-lite-body-2 and dasLOL/zils-lite-skills-v5, on vllm-sr/d3-lite / Qwen/Qwen3.5-0.8B), submitted as a body to the Zils d3-lite soup pool on Bittensor subnet 127.

Training

• Recipe: the Zils reference recipe (zils.d3 train: LoRA rank 16 on the language layers, AdamW, learning rate 2e-5, one epoch, seed 557) started from the base body above, merged into the bf16 weights with zils.d3 candidate. Every row was trained on soft target probabilities.
• Data (9,218 decision examples in the soup exam's format):
  • 4,180 tool-use cases built from Salesforce/xlam-function-calling-60k (CC-BY-4.0): counting the required arguments a request leaves missing, deciding whether a tool should be called now, and choosing between calling, asking for a detail and declining. Labels come from the dataset's verified calls; missing arguments are made by replacing a stated value with a vague reference or by making an unmentioned optional parameter required.
  • 3,580 rehearsal cases from the training splits of public NLP datasets (natural-language inference, paraphrase, similarity, topic and intent classification, question answering, function calling, commonsense, arithmetic and affect tasks), whose targets are mostly the base body's own probabilities (self-distillation) so the rest of its behaviour is kept.
  • 1,458 cases from CommonsenseQA, glaive-function-calling-v2, CLINC150 and TweetEval, with targets half the base body's probabilities and half the label.

Evaluation

Scored like the soup pool's calibrated_brier (one temperature per model fitted on 20% of the items, Brier or ranked-probability skill on the rest, families weighted by the exam's category weights) on the 714 public exam seed cases (zils-ai/base-exam-seed-v1..v3), averaged over 30 calibration splits:

Model	Score (points)	Top-1 accuracy
qhoenix/zils-lite-decision-v9 (base)	61.3	76.3%
this body	62.8	75.9%

These are development measurements, not validator results.

License

Apache-2.0, as are the base models. See LICENSE and NOTICE. Training data derived from xlam-function-calling-60k is credited to Salesforce under CC-BY-4.0.