Qwen3 support and order merge
A small synthetic experiment, not a production support model. Two independently trained LoRA adapters learn custom queue codes (RUBY=billing, JADE=login, AMBER=shipping) and SKU/quantity JSON extraction. Training templates and entity IDs are disjoint from evaluation. Intent phrases recur: this tests narrow formatting/label adaptation, not broad language generalization.
Disable thinking via enable_thinking=False in the chat template. Greedy decode, 48 new tokens. Queue scoring is exact; extraction requires only sku/string and qty/integer and exact values.
Selected on validation by worst-skill accuracy, then total; no test-set selection. Scores, all prompts and outputs, dependency versions and training evidence are in results.json. Reproduce with the bundled train.py. The experiment guide is being prepared in https://github.com/theoriclabs/letsusecompute/tree/issue-8-merge-models/posts/merge-models .
Selected merge: linear-balanced
Test results:
{
"triage": {
"correct": 48,
"total": 48
},
"extract": {
"correct": 48,
"total": 48
}
}