BurrowBrief โ€” Your data stays underground. Local AI for your everyday work.

Meet Burrow ยท Download the app ยท BurrowEssay (essays)

BurrowBrief

A LoRA adapter that makes Ternary Bonsai 2 27B write emails and replies, and fix or rewrite text, in a natural voice that keeps your facts as you gave them. It is the model behind the Email, Summarize and Edit sections of Write in Burrow, and it runs on your own PC.

It was trained mostly on emails. Summaries were not in its training data, so treat summarizing as a bonus, not a strength.

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-brief.gguf -ngl 99 -c 4096 -fa on --jinja \
  --reasoning off --temp 0.7 --top-p 0.8 --top-k 20 --presence-penalty 1.5

In Burrow, put the file at models/bonsai2/burrow-brief.gguf, choose Bonsai 2 as your model, and open Write. It shares the base with BurrowEssay and Burrow switches between them per request without reloading.

The adapter was trained and evaluated with a rules system prompt. Its main rules: keep the supplied facts exactly; invent no names, dates, quantities, motives or promises; keep uncertainty ("planned", "subject to approval"); never sign off with a name the request does not give; do not apologize for a mistake the request does not mention. Thinking is off.

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 2,321 rows (1,733 emails, 542 proofreading and editing, 46 rewrites), then DPO rounds aimed at not adding facts (153 preference pairs in the final round, 35 of them replayed from the earlier round)
Sampling temp 0.7, top-p 0.8, top-k 20, presence penalty 1.5, thinking off
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

Blind scoring by the author of 40 trap prompts (warmth creep, certainty creep, attribution, no-new-information, counting), this adapter against its own previous DPO round:

This adapter Previous round
Faithfulness (1 to 5, higher is better) 4.85 4.53
Prompts with a faithfulness error 4 14
Naturalness (1 to 5) 4.20 4.28

The faithfulness gain was significant (sign test p=0.01); the naturalness difference was within scoring noise. It is a small test, scored by one person.

Known limits

  • Email-first. Its training notes say not to treat it as a general rewrite or summary adapter. It can rewrite loosely and drop details.
  • Occasional placeholders such as "[Your Name]" can appear.
  • 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
21
GGUF
Model size
0.1B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for TyPlant/BurrowBrief-LoRA

Base model

Qwen/Qwen3.8-27B
Adapter
(8)
this model