Fine-tuning danbot

3.1 Whose slop?

I have a lot of examples of my style. There are about 1.6M words of prose I authored with my human hands on this site, for example, the the slop free ones marked clearly.

If I want to turn these into training data, I need to turn them in to slop. So I take a sample of my paragraphs, and for each one, ask an LLM to turn it into slop (“rewrite this passage to be clearer and more polished”). The output of that process is the example slop I wish to eliminate. Looking at the LoRAs out there, I imagine that a few thousand such examples might be enough to nudge the model towards my style. A smaller secondary set is generated by explicitly asking for the worst slop patterns from my own slop catalogue; that covers the long-tail mannerisms the paraphrase pipeline tends to underrepresent.

I suspect, from the literature that both sheer input length (Levy, Jacoby, and Goldberg 2024) and accumulated conversation turns (Laban et al. 2025) degrade AI output. The worst of the flows probably an LLM already gravid with context. Proposal: I generate the slop across growing sessions rather than fresh conversations, feeding a fresh, unrelated paragraph each turn and keeping each rewrite, so the examples span the range from empty- to exhausted-context. If I tag each with how full the context was, I can see where the slop gets worse and weight those examples up, or at least just see how important this factor is.

Now, which model? I use Claude a lot, so maybe I could just undo Claude? But I use other models too.

Slop seems to come in two layers. There’s a deep humus under-layer — throat-clearing, “it’s worth noting”, the rule-of-three, “It’s not X ; it’s Y”, signposting, marking key phrases in bold — that every RLHF’d model produces. Atop that is the litter inhabited by model-specific tics: GPT’s “delve” and “tapestry”, Claude’s particular hedging cadence, and so on. Attar et al. (2026) study 284 interpretable linguistic features across 27 LLMs and 10 text domains. They argue that most signals are model- or domain-specific. This suggests multi-source training might be the go.

Let us call the target here a Dan-ifier rather than a de-Claude-ifier: rewrite the input as I’d write it, whatever the source. Soto, Chen, and Andrews (2025) argues that mimicking a specific human author’s style is a strong method for removing obvious AI tics, so we have hope that this is feasible.

Moreover, I would be training a small open-source model to do this text transformation, i.e. not Claude. It will have its own tics to squish. Antoun, Sagot, and Seddah (2023) argue that small open models are easier to detect— by machine admittedly, but I assume humans also could find it easy. As such, it seems especially important to generate some of the slop using specifically my own utility base-model.

In any case I don’t want to fix on Claude, because even Claude isn’t one thing, but family of various models and version.

tl;dr We generate training slop from a plural set of LLMs.

There is at least one example of this in the wild. Paneru (2026) build a 25,140-pair AI→human corpus and fine-tune BART and Mistral-7B (QLoRA) , arguing that “using two generators rather than one reduces the risk that a trained humanizer learns to undo one model’s idiosyncratic habits rather than AI-style writing more broadly”. They also target multiple human-target domains for the same reason, which I cbf doing so I hope that part is not important. They don’t run a single-source vs. multi-source ablation, so the value of this multiple model approach is unknown. Maybe I can publish that ablation if I find time? Paneru (2026) also observe some interesting failure modes — for example we can learn the correct style transform, but overcorrect, which looks great on metrics but probably not on human inspection.

3.2 Where to train and serve

I (plan to) train on Tinker and serve on Fireworks, with local MLX as the fallback when I want it off the cloud entirely.

Tinker is Thinking Machines’ managed-training API: I write a CPU-side training loop and they run the GPUs. Its API exposes forward_backward, optim_step and sample as low-level primitives, so the same platform that runs my vanilla SFT can later shape an RL or DPO loop if I preference-tune the styler. The trained adapter downloads to Hugging Face format, which keeps out options open for local inference.

Serving the fine-tuned model, now… Fireworks deployments scale to zero and accept an uploaded LoRA which is what we want for this rarely run model. When I want it fully local I merge the adapter into the base, convert to MLX, and serve it off my own machine.¹

4.1 Scoring Dan-ness directly

We can directly estimate whether text is more slop-like or dan-like by directly training a classifier on the (slop, Dan) pairs we have, via an appropriate embedding model. Some care is needed in choosing the embedding. I am more familiar with content embedding models as made famous by ai_search. Semantic embedders, it turns out, do capture style a little bit (Icard et al. 2025) but it is not their strong suit.

However, style-embeddings are a thing (Wegmann, Schraagen, and Nguyen 2022) and seem to do okay at ignoring content. StyleDistance (Patel2025StyleDistance?) is one trained on data generated with a synthetic corpus generated in a manner very much reminiscent of this blog, which suggests we can use it here too. cf LUAR (Rivera-Soto et al. 2021) which attempts to classify authorship. As a bonus Soto et al. (2024) show that these style vectors pick out machine-generated text from human text few-shot, with no AI-detector training at all.

These things are all plug and play, so we can audition ion the standard way we evaluate classifiers.

• AnnaWegmann/Style-Embedding
• StyleDistance/styledistance
• rrivera1849/LUAR-MUD

I ran a logistic regression using each embedder as a feature extractor, scored on pairs held out by source document. StyleDistance won, at 0.78 pairwise accuracy and 0.72 AUC, ahead of LUAR (0.76 / 0.71), Wegmann’s CISR (0.73 / 0.67), and the semantic embedder baseline (0.75 / 0.62). Interestingly the semantic embedder is not terrible at pairwise accuracy, but it fails to generalize, getting a poor AUC. The quality of the StyleDistance-based classified is good enough that I probably won’t bother with more elaborate pipelines such as fine-tuning my own style encoder.

4.2 Where we are now

Phase 1 has shipped. There is a manual pair-logger (uv run ai-style-log) wired in such that every time I rewrite an AI draft into my own voice, the (before, after) diff is captured at paragraph level into a JSONL file. So even before we run a single training step, the corpus accumulates with each notebook I clean up. To make the corpus more representative, I have also stripped the automatic slop-removal pass out of the /dan-voice skill: drafts Claude generates in my voice now leak realistic AI tics, which is exactly what we want as training input. Manual cleanup is now via the separate /slop-hunter skill, invoked when I want a clean draft rather than a training-data candidate.