Meet Burrow ยท Download the app ยท BurrowBrief (emails, summaries, edits)
BurrowEssay
A LoRA adapter that makes Ternary Bonsai 2 27B write essays in a plain, human voice: mostly argumentative and reflective essays, plus other kinds. It is the essay model behind Write in Burrow, and it runs on your own PC.
This is a work-in-progress checkpoint. The adapter is unmerged and small (233 MB), so it can be replaced without changing the base model.
Use it
This is an adapter, not a full model. You need the base model, Ternary-Bonsai-2-27B-PTQ1_0.gguf, and PrismML's llama.cpp build, which Burrow bundles.
llama-server -m Ternary-Bonsai-2-27B-PTQ1_0.gguf --lora burrow-essay.gguf -ngl 99 -c 6144 -fa on --jinja \
--chat-template-kwargs "{\"reasoning_effort\":\"low\"}" --temp 0.8 --top-p 0.95 --top-k 20
In Burrow, put the file at models/bonsai2/burrow-essay.gguf, choose Bonsai 2 as your model, and open Write. Burrow loads it next to a second adapter (BurrowBrief) and switches per request without reloading the base. 32k context plus both adapters does not fit an 8 GB card, so Burrow caps Bonsai at 20k while they are installed.
System prompt it was trained with:
You are a writing assistant. Write in a natural, human voice. Follow the user's requested length, format and audience exactly. Never invent statistics, studies, quotes or sources. If the user tells you facts about themselves or gives you their draft, those are the only facts about them you may use: don't add relatives, dates, jobs, places or history they didn't give. If they give no facts about themselves and ask for a personal piece, you may write as an invented narrator, unless they ask you not to.
Thinking is on: the model writes a short plan (claim, reader, reasons, objection, length), then the essay.
Details
| Base | Ternary Bonsai 2 27B (PTQ1_0 GGUF) |
| Adapter | LoRA r16, alpha 32, f16 GGUF, 233 MB. Not merged. |
| Training | Supervised fine-tune on essays written by a hosted teacher model, then one DPO round (372 preference pairs, 2 epochs) aimed at not inventing facts about the user |
| Sampling | temp 0.8, top-p 0.95, top-k 20, thinking on |
| Tested on | RTX 4060 8 GB, about 26 tokens a second with the adapter on |
| Licence | The base model is Apache-2.0. The training data was produced with a hosted teacher model; check that provider's terms before redistributing the weights. |
What the evaluation showed
Measured by the author on their own test sets, not independently reproduced.
- Invented biography: on 78 requests of types the adapter was not trained on, a hand count found 5 outputs that invent life history, against 30 for the first version. A further 9 were borderline.
- Length: 74.6% of essays on the frozen held-out set land within the requested length, against 71.2% for the first version.
- Voice: in pairwise comparisons against the first version on personal essays, it was judged at least as human-sounding.
Known limits
- It still fails the project's strict automated check for invented facts: 36% of dev essays were flagged, against a target of 5% or less. It invents much less than the first version, not never. Read essays about you before you use them.
- A recurring move: "the strongest objection isโฆ" shows up often in argumentative essays. It is the teacher's structure, learned faithfully.
- Length misses: roughly one essay in four falls outside the requested length.
- Not trained for application essays or fiction.
- Evaluated on PrismML llama.cpp b12128; Burrow bundles build 10777. It loads and writes there, but quality was not re-measured on that build.
- Downloads last month
- 20
We're not able to determine the quantization variants.
