WikiQwen-9B

Ask anything, get a how-to article. WikiQwen-9B is Qwen3.5-9B fine-tuned to answer every message the same way: as a tidy, step-by-step how-to article in markdown, with a picture caption above each step. Hand those captions to WikiQwen-Illustrator-7B and they become illustrations.

Try it first: the WikiQwen Space runs this model and draws each step with WikiQwen-Illustrator-7B.

Round 2 (current weights): the picture captions now keep one look for every recurring person, animal and object across an article, so the pictures tell one story (see Training and Evaluation). The round-1 weights are still in this repo's commit history.

What It Does

Send it a question ("how do I keep basil alive?"), a problem ("my bike chain keeps falling off") or just "hi", and it sends back a guide. Every reply has the same shape:

• a # How to … title and a short intro
• one or more ## Method N: or ## Part N: sections, or a single ## Steps
• an [IMAGE: caption] line above each step, written for the Illustrator to draw
• steps written as **N. Bold summary.** Details…, with bullets and > **Tip:** boxes where they help
• optional ## Tips, ## Warnings and ## Things You'll Need at the end

Here is that shape, trimmed. It was written by hand to show the format, so it is not a verbatim model output:

# How to Keep Basil Alive on a Windowsill

Basil wants three things: sun, warmth and a steady drink. Get those right and it will keep you in pesto all summer.

## Part 1: Giving It the Right Spot

[IMAGE: A potted basil plant with bright green leaves sits on a sunny white windowsill, with light streaming through the glass behind it.]
**1. Pick your sunniest window.** Basil needs 6 to 8 hours of direct light a day, so a south-facing window is ideal.
- Give the pot a quarter turn every few days so the plant grows evenly.

[IMAGE: Hands pour water from a small green watering can into the soil of a potted basil plant until it drips into the saucer below.]
**2. Water when the top of the soil feels dry.** Soak the pot until water runs out of the bottom, then empty the saucer.
> **Tip:** Water in the morning so the leaves dry off before the evening chill.

## Warnings
- Don't let the pot sit in standing water. Soggy roots rot quickly.

## Things You'll Need
- A pot with drainage holes
- Potting mix
- A watering can

How to Run WikiQwen-9B

Getting your first article takes about as long as boiling the kettle. The model is about 18 GB in bf16, so one 24 GB GPU is enough.

1. Install the libraries. You need PyTorch, Transformers 5 and Accelerate. The demo Space uses transformers==5.17.0.

pip install torch "transformers>=5.14" accelerate

• Optional: pip install flash-linear-attention for faster kernels on Qwen3.5's linear-attention layers (the demo Space uses it).

2. Load the model and tokenizer. The LoRA is already merged into the weights, so it loads like any Qwen3.5 checkpoint.

3. Ask your question with no system prompt. The article habit is trained in. There was no system prompt in training, and you don't need one now.

4. Turn thinking off. Qwen3.5's chat template can open a <think> block. WikiQwen was trained with it closed, so pass enable_thinking=False. Leave it on and the model muses to itself before it starts writing.

5. Sample like the demo does. Temperature 0.7, top_p 0.9, repetition penalty 1.05 and up to 1,800 new tokens. That is room for about two methods with captions.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "devon7y/WikiQwen-9B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "How do I keep basil alive on a windowsill?"}]  # no system prompt
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1800, do_sample=True, temperature=0.7, top_p=0.9,
                     repetition_penalty=1.05)
article = tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(article)

6. Pull out the captions if you want pictures. Each [IMAGE: …] line is a ready-made prompt for WikiQwen-Illustrator-7B.

import re
captions = re.findall(r"^\[IMAGE: (.+?)\]\s*$", article, flags=re.M)

Tip: Follow-up messages work (append the reply as an assistant turn and ask again). The model was trained on single questions, though, so each answer is a fresh article, not a chatty reply.

Warnings

• Don't skip enable_thinking=False. It is the one setting that changes everything.
• Long answers can now and then repeat a step. The demo Space stops generating when the same step or caption shows up a third time, and trims an unfinished last line. Do the same if you serve it.

Things You'll Need

• A CUDA GPU with room for about 18 GB of bf16 weights plus the context
• Python with torch, transformers 5.x and accelerate
• A question (any question)

The Family

Training

• Source: about 118k how-to articles from the Kiwix March 2023 archive, with text and Creative Commons illustrations by wikiHow contributors, CC BY-NC-SA 3.0 (see Attribution and License).
• From article to chat: each article became one conversation. The user turn is a realistic message written for that article by Qwen3.6-35B-A3B. The styles range from direct questions to oblique remarks and messages that aren't how-to questions at all, which is how the model learned to answer anything with an article. The assistant turn is the article in the markdown format above, keeping at most its first two methods or parts.
• Captions: every step carries an [IMAGE: …] line. (pending) of these captions describe the step's real Creative Commons illustration, written by a vision-language model (Qwen3.6-35B-A3B). The rest were written from the step text alone. The Illustrator was trained on captions in the same style, so the chat model writes prompts the image model already understands.
• Size: (pending) training conversations and (pending) for validation. A 1% bucket of articles never entered training, and the held-out evaluation prompts below come from it.
• Round 2, consistent captions (current weights): the round-1 captions were written one picture or one step at a time, so a recurring subject often changed color or look between steps. The untuned Qwen3.6-35B-A3B rewrote all captions of 25,300 training articles together, giving each recurring person, animal or object one fixed look and quoting the exact short text a picture should show, without touching the article text. The round-1 LoRA then continued for 734 steps (learning rate 5e-5, effective batch 16) on the 9,383 rewritten articles whose consistency improved and whose captions kept their step's own details, mixed with 2,346 unchanged articles that already had consistent captions.
• Recipe (round 1): supervised fine-tuning with LoRA (rank 32, alpha 64, dropout 0.05) on the attention and MLP projections (q, k, v, o, gate, up, down). Loss was on the assistant turn only. One epoch, learning rate 1e-4 with a cosine schedule and 3% warm-up, effective batch 16, sequences up to 4,096 tokens, bf16, on one GPU. The adapter was then merged into the base weights.

Evaluation

The model wrote one reply per prompt for two prompt sets:

• Held-out prompts: user messages for articles the model never saw in training.
• Everyday prompts: 110 messages that mostly aren't how-to questions (open-assistant questions, everyday small talk and a few handwritten ones like "tell me a joke").

Replies were sampled at temperature 0.7, top_p 0.9 and repetition penalty 1.05, then checked automatically against the article format. These checks measure the shape of the answer. They don't check whether the advice is right.

Picture consistency (round 2)

38 prompts x 8 samples (304 articles), serving settings, round 1 vs round 2. Checks are automatic: a recurring subject "changes color" when two captions of one article give it different colors.

Validation loss on the round-1 set moved from 1.311 to 1.315, so round 2 kept round 1's writing.

Limitations

• The advice can be wrong. It sounds confident and looks tidy either way. Be most careful with health, safety, legal and money questions: check anything that matters with a qualified source.
• Everything is an article. "hi", "tell me a joke" and "what's 2 + 2" all get a how-to article back. That is the point, but it means this is not a general assistant.
• Knowledge is dated and general. It knows what the base model knows plus a 2023 snapshot of how-to articles. It can't browse, and it may invent product names, measurements or steps.
• Generated articles are not real articles from the source site, and no contributor wrote or reviewed them.
• Pictures can mislead. A picture can show a step wrongly, and long text inside a picture can still come out with typos.
• No extra safety training. The fine-tune only taught a format, and the article habit is strong. It may write an article for a request it should turn down, so don't rely on it to refuse.
• English only, and long answers can be cut off at the token limit.

Attribution and License

• Source credit: fine-tuned on article text and Creative Commons illustrations by wikiHow contributors, licensed CC BY-NC-SA 3.0, from the Kiwix March 2023 archive.
• This model and what it writes: CC BY-NC-SA 4.0. Section 4(b) of CC BY-NC-SA 3.0 allows an adaptation to be shared under a later version of the license with the same elements (Attribution, NonCommercial, ShareAlike). In short: credit the source, no commercial use, and share what you make under the same license.
• Base model: Qwen3.5-9B is released by the Qwen team under Apache-2.0 (license).
• Suggested credit line for articles you share: "Generated by WikiQwen-9B (CC BY-NC-SA 4.0), a model trained on text and illustrations by wikiHow contributors, CC BY-NC-SA 3.0."
• Independence: WikiQwen is an independent, non-commercial research project. It is not affiliated with or endorsed by the source site or its contributors, or by the Qwen team.

Citation

@misc{wikiqwen_9b_2026,
  title  = {WikiQwen-9B: a Qwen3.5-9B fine-tune that answers as illustrated how-to articles},
  author = {devon7y},
  year   = {2026},
  url    = {https://huggingface.co/devon7y/WikiQwen-9B}
}