This repository contains the direct merged checkpoint reproduced from wnfldchen/gemma-4-31B-it-qat-q4_0-unquantized-heretic. The original adapter repo is intentionally left intact; this repo bakes that selected Heretic edit into the base weights.

For production serving, prefer the W4A16 compressed-tensors derivative: wnfldchen/gemma-4-31B-it-qat-w4a16-ct-heretic-merged-direct.

Try the hosted product at https://chat.winfieldresearch.com.

This is a decensored version of google/gemma-4-31B-it-qat-q4_0-unquantized, made using Heretic v1.4.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation)

[!TIP] This model is reproducible!

See the README in the reproduce directory for more information. This repo was exported with export_strategy=merge, so it does not require loading a separate LoRA adapter at inference time.

Abliteration parameters

Parameter	Value
trial	6
target_components	`attn.o_proj`
ara_lora_rank	256
start_layer_index	24
end_layer_index	54
preserve_good_behavior_weight	0.1182
steer_bad_behavior_weight	0.0050
overcorrect_relative_weight	0.8715
neighbor_count	15

Performance

These are the Heretic optimization metrics for the reproduced selected trial.

Metric	This model	Original model (google/gemma-4-31B-it-qat-q4_0-unquantized)
Keywords	8/100	99/100
KL divergence	0.0900	0 (by definition)

W4A16 deployment benchmark comparison

These results compare the deployed W4A16 compressed-tensors Heretic merged artifact, wnfldchen/gemma-4-31B-it-qat-w4a16-ct-heretic-merged-direct, against the official Google W4A16 base, google/gemma-4-31B-it-qat-w4a16-ct. The benchmark artifacts are archived with the reproduction output.

Benchmark	Samples	Metric	Official W4A16 base	Heretic merged W4A16	Delta
PIQA	1,838	acc_norm	55.71%	55.01%	-0.71 pp
WinoGrande	1,267	accuracy	50.99%	50.83%	-0.16 pp
CommonsenseQA	1,221	accuracy	23.26%	23.10%	-0.16 pp
EQ-Bench	171	score	42.54	46.59	+4.05
EQ-Bench parseability	171	percent_parseable	83.04%	78.36%	-4.68 pp

PIQA, WinoGrande, and CommonsenseQA changes are smaller than their standard errors. EQ-Bench improves by 4.05 points, with lower parseability on the generative EQ-Bench scorer.

Provenance

Base model: google/gemma-4-31B-it-qat-q4_0-unquantized
Base commit: 1e4d8beecacb8b7590c1d8bedd7335f687bf311f
Source adapter: wnfldchen/gemma-4-31B-it-qat-q4_0-unquantized-heretic
Adapter commit: 8cce1a7166cd2b109886feff15ee12a78cc3177b
Heretic commit: ee47df2fbdd61d8d3a67910ea58c45f91e868917
Export method: Heretic reproduce with export_strategy=merge and model_action=save
Weight format: unquantized Hugging Face safetensors shards
Shards: 13

Load it

python

from transformers import AutoProcessor, Gemma4UnifiedForConditionalGeneration

repo_id = "wnfldchen/gemma-4-31B-it-qat-q4_0-unquantized-heretic-merged"

processor = AutoProcessor.from_pretrained(repo_id)
model = Gemma4UnifiedForConditionalGeneration.from_pretrained(
    repo_id,
    device_map="auto",
)

Notes

This is the canonical unquantized merged artifact. The W4A16 repo linked above is the smaller deployment artifact produced by compressing this checkpoint after the merge.