Humanlike Qwen3.8-27B LoRA 2.0 Adds Tool Calls, Fooled a Blind Judge 23.5% of the Time
kvyb · reddit · 2026-10-03
The author released Qwen3.8-27B-Humanlike-Chat 2.0 (v1 got 700+ upvotes, 44k downloads), keeping the human-texting voice while fixing v1's flaws: broken tool calls, ignored instructions, single default personality.
What 2.0 does
- Texts like a person with no system prompt; takes character cards; switches to formal emails/numbered steps on request; sticks to instructed habits (e.g. full sentences).
- Working tool calls — asks for missing info instead of making it up (asks your departure city when booking flights); code and math at roughly base-model level.
Training: v1 was plain SFT on 139,845 messages, which copied bad habits. 2.0 uses on-policy distillation: the model writes its own replies and two teachers grade every token — v1 with a hidden "text like a person" instruction for chat, the plain base model for instructions/tools/code. The student never sees the hidden instruction. Second LoRA merged onto the same 27B.
Numbers: IFBench 37.3→43.7, When2Call 48→58, BFCL irrelevance 60→78; MMLU-Pro and LiveCodeBench slightly worse. Self-built "ishuman" blind benchmark (judge picks which continuation was human-written): base 0.3%, base+system prompt 6.8%, official Qwen3.8 15.1%, 2.0 at 23.5% — you can't prompt your way there. 2.0 won 16/16 live multi-turn chats vs base. GGUFs and LoRA on Hugging Face.
More from Models
- Rumor: Deactivated account claims Google has far more powerful internal models than Gemini 4 Argon — mark_k · 2026-10-03
- Using System One models in Swift: fast, deterministic decisions via Apple Foundation Models — rxwei · 2026-10-03
- Linux Kernel CVEs Surge From ~500 to 1500+ Per Release, LLMs Blamed for Bulk of the Rise — burny_tech · 2026-10-03
- Developer Complains OpenAI's Coding Model Endlessly Scopes Creeps Instead of Finishing Tasks — DavidWells · 2026-10-03
- Sonnet 5 Spotted in Google Antigravity Backend, Which Still Runs Sonnet 4.6 — brandon_galang · 2026-10-03
- Post-training Yandex AliceAI-80B-A3B from scratch: a NaN bug in custom V100 kernels killed one run — jjusko20 · 2026-10-03