Bonsai-1.7B decision LoRA (training-free)

A single LoRA adapter for the Ternary-Bonsai-1.7B GGUF that raises the JevBench decision exam (231 tasks) from 124 to 143 (147 with the top3 serving ensemble).

Produced training-free: no fine-tuning, no distillation, no new data. The source decision delta is amplified ×2.5 and baked into one rank-64 file.

Quickstart

The adapter runs on the bonjev engine (https://github.com/ChezRD/bonjev), which loads the base GGUF and attaches the adapter in one step. Serve with BONJEV_LORA:

# default
BONJEV_LORA=m17_tiny25_single.gguf bonjev serve --model ternary-bonsai-1.7b --port 8830

# best score (multi-style serving ensemble)
BONJEV_STYLE=top3 BONJEV_LORA=m17_tiny25_single.gguf bonjev serve --model ternary-bonsai-1.7b --port 8830

Files

file	role	231
`m17_tiny25_single.gguf`	default (rank 64, q8_0)	143 (147 with `top3`)

Results

JevBench-231 (deterministic, BONJEV_NO_REUSE=1):

config	style	231
base	—	124
adapter	default	143
adapter	top3	147

Decision Index: not claimed for this adapter — the apparent gain is not robust (it comes from a few small tracks), so do not use it for calibrated-decision claims.

Intended use / out-of-scope

Use: typed decisions (choice / yes-no / score) with the bonjev engine.
Out of scope: free-form generation.
Limitations: the least capable size (it can fail multi-step control); the top3 ensemble is a serving mode, not part of the file.

License

Released under the Apache License 2.0 (see LICENSE). The adapter is a derivative of Apache-2.0 source models; the required attribution and notices are in NOTICE in this repository.