The maintainer of 100toolkit.com demonstrates that roughly 2,000 hand-curated input/output pairs are enough to strip the characteristic 'delve/tapestry/in the ever-evolving landscape of' vocabulary from Qwen3-4B. His core claim is that slop isn't a sign of model weakness but a surface-level artifact of RLHF annotator preferences that peels off with surprisingly little data on a consumer GPU.
The editorial highlights sample outputs where the base Qwen3-4B produces predictable ChatGPT-cadence (three-sentence intro, bulleted middle, 'In conclusion' outro) while the fine-tune generates something closer to a Slack message from a human engineer — one paragraph, no scaffolding, no apologies for brevity. This reframes the debate from 'models can't write like humans' to 'models were trained to write unlike humans.'
Commenters in the 87-comment thread pasted sample outputs and argued over whether the deslopped prose was actually better or merely trading one set of tics for another. The debate itself functions as a signal: evaluating 'human-sounding' writing is subjective and the fine-tune's success depends on which failure mode you find more tolerable.
A Show HN post landed this week from the maintainer of 100toolkit.com announcing a fine-tuned variant of Qwen3-4B trained specifically to eliminate the vocabulary and sentence-rhythm tics that make LLM output instantly identifiable. The project, which the author calls a 'deslop' model, targets the usual suspects: *delve*, *tapestry*, *navigate*, *in the ever-evolving landscape of*, *it's important to note*, *boundaries*, and the compulsive bullet-list hedging that infects almost every public model's prose. The post hit 113 points on Hacker News and generated the kind of thread that is itself a signal — commenters pasting samples, arguing about whether the output is actually better or just differently bad.
The recipe, by the author's own account, is unglamorous. Roughly 2,000 hand-curated input/output pairs where the input is a slop-laden draft and the output is the same content rewritten in direct prose. The fine-tune runs on top of Qwen3-4B — small enough to train on a single consumer GPU, large enough that the base model already has competent instruction-following. The interesting claim isn't that the fine-tune makes Qwen smarter; it's that the slop vocabulary is a thin stylistic layer that peels off with surprisingly little data.
Sample outputs in the thread back this up. Asked to summarize a technical paper, the base Qwen3-4B produces the expected ChatGPT-cadence output: three-sentence intro, bulleted middle, 'In conclusion' outro. The fine-tune produces something closer to how a human engineer would write a Slack message — one paragraph, no scaffolding, no apologies for the brevity.
The default assumption for the last two years has been that AI writing tics are a *capability* problem — the model doesn't know how humans actually write, so it defaults to the mean of its RLHF reward signal, which was shaped by annotators who liked thorough, hedged, structured answers. If that were true, you'd need a bigger model or a smarter reward model to fix it.
This project suggests the opposite: slop is a style problem that lives in a very small subspace of the weights, and 2,000 examples is enough to move it. That's a meaningful finding, because it inverts how people have been thinking about model personality. The frontier labs have been treating voice as something that emerges from scale and alignment. A LoRA fine-tune on a 4B-parameter model is challenging that with a much cheaper hypothesis: voice is a cheap post-training patch, and the only reason every model sounds the same is that every lab is optimizing against similar annotator preferences.
The second-order implication is more interesting. The entire 'AI detector' industry — Turnitin, GPTZero, every university's panicked policy document — is built on the assumption that LLM output has consistent, detectable stylistic fingerprints. If a single developer can train those fingerprints out with 2,000 examples on a $0.50/hour GPU, the detector business has a problem that no amount of statistical analysis will solve. The fingerprints aren't a law of nature. They're an artifact of training-data overlap between the major labs.
Community reaction split along predictable lines. The ML practitioners in the thread were mostly impressed by how little data moved the needle — one commenter noted this matches recent papers showing that style transfer in LLMs is cheaper than content transfer by roughly an order of magnitude. The skeptics pointed out, fairly, that eliminating one set of tics usually just introduces another. If the training data is 2,000 examples from one author, the output will start sounding like that author. You've traded ChatGPT voice for whoever-wrote-the-training-set voice.
There's also a real question about whether the fine-tune degrades capability. The author claims it doesn't, but the HN thread contains no systematic eval. On a 4B model, you'd expect some capability tax from any fine-tune — the question is whether it's visible on tasks the user cares about. Nobody has run MMLU or HumanEval on the deslop variant yet, which is both the correct engineering move (ship first, benchmark later) and a reasonable thing to be suspicious of.
If you're building anything that generates user-facing text — docs, summaries, email drafts, support responses — this project is worth 30 minutes of your attention. Not because you should use this specific model, but because it demonstrates that the 'make it sound less like AI' problem is solvable at the fine-tune layer rather than requiring a prompt-engineering arms race. Most teams right now are spending prompt tokens on instructions like 'do not use the word delve' and 'avoid starting sentences with In the.' That's a losing game — you're paying inference cost on every request to fight a problem that could be baked into the weights once.
The practical move, if you have the data: curate a few thousand examples of your actual house voice (your docs, your support replies, your changelog entries) and run a LoRA fine-tune on a small open model. The 100toolkit author did it on Qwen3-4B; the same approach works on Llama 3.1 8B, Mistral Small, or Gemma 2 9B. Training cost is in the single-digit-dollars range. You'll end up with a model that writes like your company instead of like the median RLHF annotator, and you'll stop burning prompt tokens on negative instructions.
The caveat for enterprise teams: fine-tuning changes your eval surface. A model that's been style-tuned on your corpus will have subtly different failure modes than the base, and your regression suite needs to catch stylistic drift as well as factual errors. If you're already running evals on a hosted model, moving to a fine-tuned local one means owning that eval pipeline yourself. For a lot of teams that's a feature, not a bug — but it's work.
The thing to watch isn't this specific model. It's whether the pattern generalizes. If 2,000 examples can strip slop from a 4B model, what else is a thin stylistic layer? Formality? Hedging? The compulsive 'Great question!' preamble? There's a plausible near-term future where every serious product team maintains an internal LoRA for voice the same way they maintain a style guide — and where 'sounds like ChatGPT' becomes a tell that you shipped on defaults instead of doing the work.
Why kill the em dashes? They are a legitimate mark — one that I have been using for years. On a Mac, you quickly get used to pressing shift-option-hyphen. Em dashes are typographically correct and should be used.
In before: Goomba fallacy. But I can't notice how there's claims that: LLMs are nondeterministic and cannot be relied upon to produce the same thing AND they produce the same texts, same UIs, in the same styles. It frustrates me too, but what a solution for both problems would look like?Al
I wonder if dynamically changing the temp based on the requested tone + a small "how to write skill" could produce more human varied writing, instead of using the same temp for the full request.
and needs no system prompt A human would write "and doesn't need a system prompt."
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Models have proclivities, and many people use the same model.Imagine all books on earth were written by just 5 people - and worse, they could not remember what they had written yesterday. You would quickly get tired of reading the same flourishes and patterns over and over.Whatever style models curr