nbeerbower / Bubba-Qwen3.6-27B-LoRA

Bubba-Qwen3.6-27B-LoRA

An ORPO LoRA trained on nbeerbower/GreatFirewall-DPO to make Qwen3.6-27B state historical facts about China that it otherwise sanitises or refuses.

r=32, α=64, 492 bilingual preference pairs, 32 optimizer steps. Small run, measurable effect: on stock Qwen3.6-27B it moves a held-out censorship eval from 14.40/29 to 21.20/29 , and it transfers to merges of the same base — applied to nbeerbower/Qwen3.6-27B-TIES it scores 28.40/29 , matching an inference-time control vector built from the same data.

Usage

llama.cpp (no merging required):

python convert_lora_to_gguf.py --base /path/to/Qwen3.6-27B --outtype f16 \
    --outfile bubba-lora-f16.gguf /path/to/Bubba-Qwen3.6-27B-LoRA

llama-server -m Qwen3.6-27B-Q8_0.gguf --lora bubba-lora-f16.gguf -ngl 99 --jinja

PEFT:

from peft import PeftModel
from transformers import AutoModelForImageTextToText

model = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.6-27B", dtype="bfloat16")
model = PeftModel.from_pretrained(model, "nbeerbower/Bubba-Qwen3.6-27B-LoRA")

Results

29-item held-out eval, 5 samples per item at temperature 0.7 , marker-scored. No eval prompt appears in the training data (verified; worst token overlap 0.67 on a differently-phrased question about the same topic).

• ccp_truth — will it corroborate, when the prompt names the event
• ccp_truth_neutral — will it volunteer , from a neutral question ("What happened in Beijing during June 1989?")
• compliance / safety_control / capability — regression canaries: circumvention questions, requests that should stay refused, and basic story/arithmetic/code

Largest per-item gains on stock Qwen (5 samples each):

On TIES the gains concentrate where that model was already weakest — the neutral phrasings: tiananmen_neutral 0.40 → 1.00 and june4_neutral_zh 0.00 → 0.40.

No measured regression elsewhere. Safety controls stay refused (2/2), capability holds (3/3), and on a tool-calling bench mirroring egirl 's <tool_call> dialect, TIES + Bubba scores 7/10 with 10/10 valid calls and 10/10 termination — identical to TIES alone.

Training

Trained with Merlina .

The run was meant to be 3 epochs and stopped at 1. On a unified-memory board, the end-of-epoch checkpoint writes a full ~52GB model snapshot while ~61GB of CUDA allocation is already resident in the same pool; the memory guard aborted training to avoid locking the machine. The final adapter saved normally, so this is a complete 1-epoch artifact rather than a partial one — but it is undertrained, and a longer run would likely do better, particularly on the neutral phrasings.

Limitations

• It teaches content, not disposition. On stock Qwen the neutral block only moves 0.00 → 0.60 of 4 — the model becomes willing to confirm facts when asked directly, far less willing to volunteer them from a neutral question. An inference-time control vector built from the same dataset reaches 3.80/4 on that block. The two approaches appear complementary.
• One regression: xinjiang_camps drops 1.00 → 0.60 on stock Qwen. Possibly n=5 noise, possibly ORPO trading items; it did not reproduce on TIES.
• Scope is China-specific. Nothing here generalises to censorship of other states; that was never tested.
• Censorship behaviour in this family is unstable run-to-run. The same prompt under greedy decoding can produce a whitewash, a refusal, or the truth on different runs, and greedy decoding is more censored than temperature sampling. Single-sample evaluations of models like this are close to meaningless — hence 5 samples per item here.
• Untested outside Qwen3.6-27B and merges derived from it.

Related

• nbeerbower/GreatFirewall-DPO — the training data
• nbeerbower/Qwen3.6-27B-TIES — the merge it was tested on