noonghunnaandClaude Opus 4.7
84498d47aa
Release / release (push) Failing after 50s
Vendored snapshot of froggeric/Qwen-Fixed-Chat-Templates qwen3.6
template, mounted into all 22 vanilla Qwen 3.6-27B composes via
--chat-template. Replaces the model's default Jinja template with
the community-patched one. Carnice and Qwopus composes intentionally
excluded — they ship bespoke Hermes-JSON templates that must not be
overwritten.
Upstream fixes seven documented bugs in the default Qwen 3.5 / 3.6
templates:
- empty <think></think> blocks polluting past-turn context
- </thinking> closing-tag hallucination on Qwen 3.6
- unclosed <think> before tool_call (mangled output)
- raise_exception crash when no user query in messages (kills
agentic loops)
- "developer" role rejection (blocks modern API clients)
- |items Jinja filter unsupported in C++ runtimes (llama.cpp,
LM Studio, MLX)
- type-aware tojson serialization for tool arguments
Source: https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
Surfaced by @troymroberts in club-3090 discussion #121.
First-pass A/B 2026-05-12 vs the matched-config Qwen INT8 PTH n=4
rebench baseline (2026-05-10):
Pack Baseline Froggeric Δ
---------------- -------- --------- --
toolcall-15 10/15 10/15 0
instructfollow 13/15 13/15 0
structoutput 13/15 13/15 0
dataextract 15/15 15/15 0
reasonmath 6/15 6/15 0
bugfind 11/15 11/15 0
hermesagent-20 9/20 12/20 +3 (+15pp)
cli-40 17/40 17/40 0
TOTAL 94/150 97/150 +3 (+2pp)
hermesagent-20 is the multi-turn agentic pack — exactly where the
empty-think + no-user-query + unclosed-think-before-tool-call fixes
compound. 7 other packs flat = no regression on single-turn flows.
Added a new "Community templates / model assets" section to
docs/UPSTREAM.md tracking the resource + drop trigger (replace when
upstream Qwen pushes equivalent fixes).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>