license: apache-2.0
base_model: Qwen/Qwen3.5-4B
library_name: mlx
pipeline_tag: text-generation
tags:
- mlx
- bf16
- experimental
Qwen3.5-4B — weight-edited attempt 1 (bf16)
Experimental research checkpoint. This is the existing historical attempt 1, selected on validation data,
not a newly selected direction or a claim of a fully refusal-free model. The original base weights are
preserved separately. Modified weights are supplied in the source-style safetensors layout, tested for
text chat with mlx-lm 0.32.2 and MLX 0.32.4 on Apple silicon.
The edit projects a selected direction out of decoder residual writers, embeddings and the vision merger.
The tied output read path changes with the embedding. The original export omitted MTP tensors. Image
behavior and MTP-assisted generation are not verified.
Existing saved measurements
| Variant | JBB marker refusals / 100 | First-token harmless KL | GSM8K accuracy |
|---|---|---|---|
| Original bf16 | 95 | 0 | 87.72% |
| This attempt 1 | 10 | 0.0640 | 85.90% |
The fixed acceptance bar was at most 5 JBB refusals, so this checkpoint did not pass the full bar.
GSM8K's paired change was -1.82 percentage points (SE 0.77). One benchmark does not establish general
capability preservation, and marker absence does not establish semantic compliance. These are historical
saved results; the upload did not rerun behavioral benchmarks. A local load/chat smoke test passed.
Provenance and integrity
abl_edit.json is the original edit metadata. release_manifest.json contains SHA-256 hashes of this
release. Historical source revision and extraction lineage were not completely recorded; the current
cached base snapshot must not be treated as proof of the old export's exact lineage. This historical
checkpoint has not been retrospectively certified by the new pipeline's full tensor round-trip checks.
The accompanying license is the upstream Apache 2.0 license. This repository contains modified weights;
it is not an official Qwen release. No evaluation prompts or raw harmful generations are included.
Loading
Download this public repo, then pass the downloaded folder to mlx_lm.load.
In persona-lab, use python evil_chat_mlx.py --model /path/to/download --dose 0 --max_tokens 4096.
The zero dose disables extra neuron steering; the weight edit is already in the checkpoint.
Modification and distribution notice
The weight files were modified by persona-lab using the intervention recorded in the export metadata.
Each safetensors file carries a modification notice in its header. This release packaging changes
header metadata only: every tensor's dtype, shape, offsets and payload bytes remain the same as the
verified edited checkpoint. The original local checkpoint is not rewritten.
Upstream licenses, attribution and usage restrictions continue to apply. See the included license,
notice and policy files. No benchmark prompts, raw harmful completions or manual-test programs are included.