Repository navigation
Qwen3.5-0.8B thinking-ON leg: Qwen sampling over N seeds + a required closure rate instead of greedy (GATE CHANGE) #4696
Description
GATE CHANGE: mint only. Nobody starts this until the operator signs it off in a batch (C297 §5, §7).
Proposed by aprender-ont244 (cop inbox, 2026-10-04 08:06Z PROPOSE-TICKET), verbatim:
thinking-ON leg for Qwen3.5-0.8B: replace greedy with Qwen sampling (T0.6 k20 p0.95, N seeds) + required closure rate. GATE CHANGE, so it needs operator batched sign-off. Basis handoff/c298-4663-4666-evidence.md: llama closes 2/8 (Q4_K_M) and 4/8 (IQ4_XS); apr Q4_K_M 5/8
Basis (cop-inbox handoff/c298-4663-4666-evidence.md, released apr v0.70.1 cuda on lambda RTX 4090, llama.cpp b11382):
- apr greedy closes the think block at no budget (2048 / 4096 / 8192 / 16384 tokens) on IQ4_XS or Q4_K_M; output grows linearly and loops.
- Greedy diverges at an early near-tie across apr-GPU, apr-CPU, llama-GPU and llama-CPU; llama's greedy runs happen to close, apr's loop.
- With Qwen's recommended sampling (T 0.6, top-k 20, top-p 0.95), 8 seeds, 2048 tokens: Q4_K_M llama 2/8, apr 5/8; IQ4_XS llama 4/8.
- The evidence reads it as a model property of Qwen3.5-0.8B in thinking mode, not an apr forward defect.
Relates to #4663 and #4666. Decision needed from the operator: change the leg as above (N and the required closure rate to be set), or keep greedy / refuse the model.
P1, under the 0.70.2 epic. Owner: aprender-ont244. Minted by the cop's clerk.
Activity
fix(qa): the #4696 sampled thinking-ON leg reaches the hybrid runtime…
noahgift commentedon Oct 8, 2026
Parked by owner aprender-a7n (moved to a release-script defect). State for whoever picks it up:
build-kaizen/4696-sampled-think-on-r2@ e920d8d holds the original two commits, re-cut ontomain. Thethinking-budgets-v1.yamlconflict (both copies) keeps main's 0.70.2 P0: qwen35-0.8b-q4km golden_output thinking_on: think block unclosed within 2048 tokens (known failure in 0.70.1) #4666 basis text and adds the sampling block.pv validatepasses on both contract copies. The ladder self-test has 156 cases and 0 bad, including the sampled green row and the fault and greedy-text red rows. The apr-cli golden/thinking/output_verification tests pass 75/75, and the aprender-contracts lib tests pass 2290.Merge origin/main into build-kaizen/4696-sampled-think-on-r2 (#4696)
fix(ladder): #4696 retires the qwen35-0.8b-q4km known_red instead of …
noahgift commentedon Oct 14, 2026
merged into #5317 (master ticket for epic #4381; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.