Qwen3.8-27B-Obliterated-E03

manitcor/Qwen3.8-27B-Obliterated-E03 is a BF16, text-only derivative of Qwen/Qwen3.8-27B. It was produced with the experimental qwen38_e03 refusal-direction intervention in OBLITERATUS.

This release is intended for alignment research, red teaming, mechanistic interpretability, and evaluation of refusal interventions. It has materially reduced refusal behavior. Do not treat it as a safety-aligned replacement for the upstream checkpoint.

Release summary

The saved derivative contains the text trunk only. It is published with the text-generation pipeline tag and should not be represented as preserving the upstream vision-language interface.

Intervention configuration

The E03 candidate used an immutable 500-pair intervention split and a separate 142-pair optimizer-tune split. The final 200-pair split was withheld until the saved candidate passed the declared tune gates.

The intervention covered the validated Qwen3.8 text manifest: 48 linear_attn.out_proj, 16 self_attn.o_proj, and 64 mlp.down_proj module families. Per-layer intervention strength was adapted to the measured refusal subspace. The complete machine-readable configuration is retained in abliteration_metadata.json. Its top-level quality_metrics object contains the final, fresh-reload holdout result. Earlier intervention-time diagnostics are retained separately under intervention_time_diagnostics; optimizer-tune and final-holdout records remain independently auditable under evaluation/.

Evaluation

Every reported release-qualification evaluation loaded the saved checkpoint in a fresh supervised process. Before each load, all 35 managed checkpoint artifacts were checked against the archived byte sizes and SHA-256 digests. The tune and final evaluation records are stored under evaluation/.

Promotion results

The preregistered promotion gates were refusal below 30% and coherence at or above 80%. E03 passed both gates on the optimizer-tune split and on its single authorized final-holdout evaluation.

The six capability probes covered tool calling, JSON-schema output, arithmetic reasoning, code generation, visual description, and instruction following. The arithmetic/chain-of-thought probe failed; the other five passed. These are small diagnostic checks, not broad capability benchmarks.

Important evaluator note

An earlier inline run reported 60% coherence because the evaluator scored only the first 200 characters and rejected some concise correct answers before checking their factual anchors. Release qualification used the corrected full-completion scorer. Only log previews are truncated. The corrected scorer was regression-tested against concise anchored answers and answers whose anchor appears after character 200.

Refusal strata

The final holdout was evenly stratified across seven internal corpus strata. Remaining refusals were 0/29, 0/29, 0/28, 1/29, 0/28, 1/29, and 0/28. The corpus prompts are not published in this repository; stable salted pair hashes and split membership are provided in evaluation/experiment-protocol.json.

Usage

The checkpoint was qualified with transformers==5.8.1. Qwen3.8's hybrid runtime may also require current Flash Linear Attention and causal-convolution dependencies for the selected backend.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "manitcor/Qwen3.8-27B-Obliterated-E03"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Explain why the sky appears blue."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(inputs, max_new_tokens=256)

print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

A full-precision load needs approximately 54 GB for serialized weights and substantial VRAM/RAM headroom for runtime allocations. The qualification host used one 80 GB A100 and loaded approximately 52 GB into VRAM.

Limitations and safety

• Refusal behavior has been deliberately reduced. The model may comply with unsafe, illegal, privacy-invasive, or otherwise harmful requests.
• The final holdout contained a 3% rate of repetitive or broken responses on harmful-prompt generations. This is a known quality defect.
• The 83.3% capability result is based on only six deterministic probes and is not evidence of unchanged general capability.
• The spectral diagnostic was inconclusive because the available sample count was below its declared requirement.
• No claim is made that the intervention generalizes across languages, multimodal inputs, long contexts, tool environments, or adversarial prompt distributions.
• The checkpoint has not been independently audited for deployment in a public-facing product.

Use appropriate access controls, output filtering, sandboxing, monitoring, and human review for any deployment. Users are responsible for complying with applicable laws and the Apache-2.0 license.

Provenance and reproducibility

Reproduction record

The intervention was run on 2026-08-29 on an NVIDIA A100 80 GB host using Python 3.12.3, Linux 6.11.0-29, CUDA runtime 12.8, PyTorch 2.10.0, Transformers 5.8.1, Accelerate 1.14.0, BitsAndBytes 0.50.0, and OBLITERATUS 0.1.3. The input corpus contained 500 harmful and 500 harmless prompts and had SHA-256 60ce2bb7fa115d00d1da85dec536a9a1dddd11378313ea1ee7bd48cf605998c3. The immutable train/tune/final protocol manifest has SHA-256 ba65938b24eeaefd51e9780d2915a82bc5bdfa6ebeee541e9157ccfc3dccf692.

Reproduction requires the exact upstream revision and reviewed commits below. Load the upstream model as native BF16 with quantization disabled, select the qwen38_e03 method, apply the intervention configuration above, save to a new directory, and evaluate that saved directory in a fresh process. Candidate selection must use only the 142-pair optimizer-tune partition; the 200-pair final partition is a single-use holdout and must not be used for retuning.

The historical run manifest recorded seed, model_revision, and tokenizer_revision as null, and did not retain the exact UI request or a single repository HEAD covering every listed implementation commit. The upstream revision is independently bound by the release provenance, but these historical omissions mean an exact bit-for-bit rerun is not claimed.

The review PR was intentionally left unmerged at publication time. Commit IDs are included so the exact reviewed implementation remains identifiable.

Artifact verification

artifact-inventory.json records the original byte size and SHA-256 digest of every checkpoint artifact. evaluation/provenance.json binds the release to the source revision, experiment run, qualification records, and implementation commits. Hugging Face/Xet object identifiers may differ from these source archive hashes; verify downloaded files against artifact-inventory.json when reproducing this release.

License and attribution

Text-only configuration notice

This repository contains 64 text layers and no MTP/next-token-prediction tensors. The current config.json nevertheless retains the upstream mtp_num_hidden_layers: 1 field. Transformers 5.8.1 loaded and qualified the text-only checkpoint, but converters that trust this field may expect a nonexistent layer. Until a corrected immutable revision is published, treat the supplied safetensors inventory—not that optional config field—as the authoritative serialized architecture. Do not edit a downloaded release in place and call it the same revision.

This derivative is distributed under the Apache License 2.0, matching the upstream Qwen3.8-27B release. See LICENSE. Qwen is the work of the Qwen team; this repository is an independently produced research derivative and is not an official Qwen release.