license: apache-2.0
base_model: openai/gpt-oss-20b
library_name: transformers
pipeline_tag: text-generation
tags:
- bf16
- experimental
GPT-OSS-20B-abliterated-dequantized-bf16
Behavioral status — 2026-10-31
Refusal removal has not been demonstrated. Manual chat testing reported continued refusals on the harmful test requests tried. This checkpoint should not be described as successful refusal removal. The observation is a user-reported spot check, with no matched base-model run or quantified refusal rate.
All saved tensor checks, fresh-process reload checks and a benign chat smoke passed. Those establish storage/inference integrity. The strict projection diagnostic was 1.0758% against a 1% target, and failed that reference threshold. Held-out harmless first-token KL was 0.026350. Broad capability preservation remains untested.
The saved protocol scores first-token changes after an ordinary analysis-channel opener. Its relationship to final-answer refusal labels has not been validated for GPT-OSS. No raw harmful prompts or completions are included here.
Public experimental weight-edited checkpoint derived from openai/gpt-oss-20b.
This independent modification is not an official release from the original publisher. Original weights
were SHA-256 checked before and after export; edited tensor bytes, shapes and dtypes were checked after
saving. A fresh-process reload matched an edited in-memory reference on three fixed harmless prompts
over three uncached greedy steps. These checks establish serialization integrity.
Behavioral refusal removal and broad capability preservation have not been established. The checkpoint
can still refuse requests; one conversation or a phrase-counting refusal score is insufficient evidence.
The saved intervention policy, direction, selection, data/code hashes, environment and checks are in .pipeline/.
Image behavior, long-context behavior and MTP-assisted generation are untested.
Export diagnostics
Held-out harmless first-token KL (100 prompts): 0.026350. Maximum edited/base projection ratio: 1.076%. The 1% mechanical diagnostic fails. These are diagnostic references, not complete behavioral/capability acceptance.
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.
Precision provenance and chat
The publisher released the expert weights in MXFP4. This checkpoint was edited from a separately saved, verified bf16 dequantization of that release. Dequantization does not recover precision lost during quantization; this is not an original full-precision OpenAI release. The unedited comparison baseline is that same dequantized representation. Conversion provenance is retained in .conversion/ and the source manifest. Use the separate requirements/native-cuda.txt environment and chat_cuda.py, which uses eager attention with sinks, native eager experts, and the Harmony template at low reasoning effort with an explicitly opened ordinary analysis channel. This same prompt policy is used for extraction, original/edited checks and chat, so the first-token proxy scores content instead of a fixed protocol header. The terminal displays only the final assistant channel. No tools are executed.