How it was trained
All training data is permissively licensed, so the model is released under Apache 2.0.
SFT (LoRA r=128 on all linear layers, merged, 2 epochs, Unsloth, padding-free, 8× RTX 5090, 28 minutes), 23k examples:
- Human replies: the best-ranked human assistant replies from OpenAssistant oasst2
(Apache-2.0), English only, quality- and toxicity-filtered, one example per assistant turn.
- Human long-form writing: documents from fineweb-edu
(ODC-BY) crawled in 2019–2020, before AI-generated web text was common. The base model wrote the request each
document answers (instruction back-translation); the response is the untouched human text.
- Capability replay: the base model's own answers to chat (UltraChat, MIT), math (GSM8K/MATH, MIT) and code
(Magicoder OSS-Instruct, MIT) prompts. Math and code keep their full reasoning in thinking mode, so the model
doesn't drift away from what it could already do.
A DPO stage (human reply preferred over the base model's reply) was also trained, but it is not part of this
release: it lowered non-thinking GSM8K from 90.5% to 82.8% and HumanEval from 78.0% to 68.3%.
Evaluation
All prompts are held out from training. Sampling for every model: temperature 0.7, top-p 0.8, top-k 20, non-thinking mode.
HumanQwen long-form uses presence_penalty=1.5 (see Usage).
Style. Chat: 300 held-out oasst2 conversations, compared with the real human reply. Long-form: 200 held-out writing
requests, compared with the real 2019–2020 human document.
Table with columns: Chat replies, Qwen3.5-4B, HumanQwen3.5-4B, Real humans| Chat replies | Qwen3.5-4B | HumanQwen3.5-4B | Real humans |
|---|
| Length (words) | 549 | 209 | 179 |
| Replies with markdown headers | 79% | 6% | 4% |
| Bold phrases per 100 words | 3.6 | 0.24 | 0.09 |
| Lines that are bullet points | 40% | 26% | 22% |
Table with columns: Long-form writing, Qwen3.5-4B, HumanQwen3.5-4B, Real humans| Long-form writing | Qwen3.5-4B | HumanQwen3.5-4B | Real humans |
|---|
| Length (words) | 628 | 507 | 589 |
| Pieces with markdown headers | 85% | 0% | 0% |
| Bold phrases per 100 words | 1.77 | 0 | 0 |
| Lines that are bullet points | 23% | 8.5% | 5.7% |
| Stock LLM phrases per 1k words ("delve", "it's important to note", …) | 1.35 | 0.86 | 0.33 |
Capabilities (GSM8K: 400 test problems; HumanEval: all 164, executed):
Table with columns: Qwen3.5-4B, HumanQwen3.5-4B | Qwen3.5-4B | HumanQwen3.5-4B |
|---|
| GSM8K, thinking | 85.8% | 87.0% |
| HumanEval, thinking | 90.2% | 90.9% |
| GSM8K, non-thinking | 91.5% | 90.5% |
| HumanEval, non-thinking | 82.9% | 78.0% |
An honest caveat about AI detectors. The open detector
fakespot-ai/roberta-base-ai-text-detection-v1
still flags HumanQwen's long-form writing as AI-generated (100% of pieces, vs. 98.5% for the base model and 15.5% for the
real human documents). The model writes with human structure and register, but not with human-like token statistics.
(On chat the detector is uninformative: it flags 91% of the real human oasst replies as AI.)
Usage
Writing style was trained in non-thinking mode (enable_thinking=False); thinking mode is kept for math and code.
Use Qwen3.5's recommended non-thinking sampling including presence_penalty=1.5. Without it, about 40% of long
pieces fall into repetition loops; with it, about 5% do.
vllm serve Thermostatic/HumanQwen3.5-4B --max-model-len 32768
from openai import OpenAIclient = OpenAI(base_url="http://localhost:8000/v1", api_key="x")r = client.chat.completions.create( model="Thermostatic/HumanQwen3.5-4B", messages=[{"role": "user", "content": "Write a short blog post about learning to bake sourdough."}], temperature=0.7, top_p=0.8, presence_penalty=1.5, extra_body={"top_k": 20, "chat_template_kwargs": {"enable_thinking": False}},)print(r.choices[0].message.content)
Limitations
- The human-written training replies come from crowd volunteers (oasst2) and 2019–2020 web pages, with their quirks.
The model can be terser than you might want.
- It reads more human, but it does not evade AI-text detectors (see Evaluation).
- Non-thinking HumanEval is about 5 points lower than the base model; use thinking mode for code.
- Long outputs can loop without
presence_penalty.
- Only English writing was targeted.
License
Apache 2.0. Base model: Qwen3.5-4B by the Qwen team (Apache 2.0). Training data: OpenAssistant oasst2 (Apache-2.0),
fineweb-edu (ODC-BY), UltraChat 200k (MIT), GSM8K (MIT), MATH (MIT), Magicoder OSS-Instruct (MIT).