Repository navigation
T1-R1: aprender-train refuses hybrid/GDN (qwen3.5) arch until modelled; fix qwen3_5 head_dim override #4552
Closed as not planned
Description
From la-72 (L2, epic #4000 'Train What You Serve'). Finding: training verbs have no qwen35 refusal; aprender-train silently models qwen3.5 as dense qwen3 (no GDN; head_dim forced to 128).
Falsifier: a qwen35 .apr into finetune/distill exits non-zero with a named refusal, until GDN fwd+bwd lands (T1).
Sub-issue of #4000.
Activity
fix(contracts): each TAH row cites its tests, not the prose after them
fix(contracts): each TAH row cites its tests, not the prose after them
noahgift commentedon Oct 14, 2026
merged into #5314 (master ticket for epic #4000; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.
noahgift commentedon Oct 14, 2026
Moved: this sub-ticket is now a row in #5341 (apr finetune and LoRA adapters: correct loading, merging and CUDA training), split out of #5314 by theme.
fix(train): refuse qwen3.5 on the distill --config train path and the…
fix(apr-cli): distill refuses an unmodelled arch on every input form (#…
fix(apr-cli): distill --config --plan/precompute and the CUDA --plan …
fix(apr-cli): finetune refuses a hybrid behind --model-size, a GGUF a…
fix(apr-cli): positional distill refuses a remote id by name; --model…
fix(apr-cli): an .apr's hf_architecture/hf_model_type stamp is a clai…