license: mit
library_name: transformers
pipeline_tag: text-generation
base_model: deepseek-ai/DeepSeek-V4-Flash-0731
base_model_relation: finetune
tags:
- deepseek-v4
- mixture-of-experts
- abliterated
- fp8
- model-surgery
- mechanistic-interpretability
DeepSeek-V4-Flash-0731 Abliterated (Native FP8)
Experimental model-surgery release. Direct behavioral benchmarking of
this native-FP8 checkpoint has not been run. The completed behavioral
validation below uses a quantized MLX derivative as a deployment proxy and
must not be read as a direct FP8 or capability result.
This is a rank-1 abliterated derivative ofdeepseek-ai/DeepSeek-V4-Flash-0731.
It remains the official post-trained 0731 model: abliteration does not turn it
into a base model, and this release is not claimed to be universally
"uncensored."
Hugging Face labels the parent relationship as finetune because its model
tree currently has no generic derived or model-surgery relationship. This
checkpoint was not gradient-finetuned and was not quantized by this project;
it was produced by a targeted projection edit while retaining the upstream
checkpoint's native mixed precision.
Provenance
- Upstream repository:
deepseek-ai/DeepSeek-V4-Flash-0731. - Exact upstream revision:
9e165c30e2704aec5d9d593cce3eebd58bbef1cb. - Refusal-direction source revision:
85ae5bbaa7daf4da7b40e007a71a1c06b41fc053. - Reproduction tooling snapshot:
apetersson/deepseek-model-tools@c3979d5, usingscripts/abliterate_dsv4.py.
Edit recipe
- Method: refusal-direction projection from attention residual writers.
- Direction rank: 1.
- Strength: lambda 3.5.
- Main-model layers: 10 through 42, inclusive.
- DSpark stages: the corresponding attention output projection in all three
attached stages was edited. - Total edited tensors: 36.
- Storage: the checkpoint's original mixed FP8/BF16/F32 representation, with
three fixed-point FP8 requantization iterations. - Direction SHA-256:
6e4d8a8f3aa9e21795faab2c5b14d29b019acdf2ddbfbd8238430458a5837fe0.
The recipe follows the public refusal-direction work indrowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-1M-57toks,
transferred to the newer 0731 checkpoint. That transfer is experimental. SeeABLITERATION_MANIFEST.json for the per-tensor edit and FP8 round-trip data.
Format and use
The architecture, tokenizer, official message encoding, one-million-token
context declaration, and attached DSpark tensors come from the upstream 0731
checkpoint. Refer to the upstream model card and the included encoding
directory for prompt formatting and runtime instructions.
Recommended upstream sampling defaults are temperature=1.0, top_p=0.95
for agentic scenarios, and top_p=1.0 otherwise.
Validation status
Structural validation
- All 48 weight shards and 72,317 indexed tensors passed structural validation.
- All 36 intended residual-writer edits are recorded in the manifest.
- Preliminary direct refusal probes showed the expected behavioral shift.
Behavioral deployment proxy (2026-08-01)
The native-FP8 checkpoint was not loaded for the full benchmark on the 128 GB
test host. Instead, the complete prompt gauntlet was run through oMLX usingapetersson/DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64,
a quantized derivative of this checkpoint. This is useful evidence that the
edited behavior survives that conversion and runtime, but it is not a direct
measurement of this FP8 artifact.
Run configuration:
- oMLX 0.5.4rc1, OpenAI-compatible API, 32,768-token profile.
- Temperature 0, top-p 1, seed 42, and maximum 160 generated tokens.
- Benchmark tooling and pinned source revisions:
apetersson/deepseek-model-tools@dc6af88. - Benchmark fingerprint:
40c3573bd48861b846721220fa06ce0d71905aab236ce45d49d9aa0e95e79af5. - Final completeness gate: 830 unique cases, 830 successful latest records,
and zero failed or missing latest records. Seven transient oMLX transport or
memory-guard failures were repaired with a serial retry pass.
| Suite | Cases | Completed result |
|---|---|---|
| UncensorBench | 200 | 0 hard refusals; keyword compliance on 200/200 |
| XSTest | 450 | 0/450 refusals under the exact upstream prefix classifier; 0/250 safe-prompt over-refusals |
| StrongREJECT-small | 180 | 60 prompts × baseline, ROT13, and refusal-suppression variants; 0 hard refusals under the local heuristic |
One XSTest unsafe privacy prompt triggered the deliberately broader local
hard-refusal heuristic while not matching XSTest's official prefix classifier;
no safe prompt triggered either classifier.
The official fine-tuned StrongREJECT judge was not run, so this card does
not claim a StrongREJECT score. A clean reference run and capability benchmark
were also not run. Consequently, these results do not establish the native-FP8
checkpoint's refusal rate, quantify the effect of quantization, or demonstrate
general capability preservation.
Limitations
Abliteration can affect capabilities and behavior beyond refusals. It does not
guarantee compliance, factuality, safety, or a particular response style. Use
appropriate access controls and evaluate the model for your deployment.
License and attribution
The upstream repository and weights are MIT licensed. This derivative retains
the upstream LICENSE. The refusal direction is attributed to drowzeys/keys
under its accompanying MIT notice; see NOTICE. Please cite the original
DeepSeek-V4 work and credit DeepSeek-AI when redistributing or publishing
results.