license: apache-2.0
tags:

• merge
• lora
• abliterated
• uncensored
• reasoning
• coding
• token-efficient
• qwen3_8
• ai-red-team
• conversational
  base_model:
• orcarouter/Qwen3.8-27B-Uncensored

Qwen3.8-27B-Uncensored-Terse-Coder

Merged weights of orcarouter/Qwen3.8-27B-Uncensored
with the Shockem/Qwen3.8-27b-Terse-Coder-LoRA
adapter (round 8, rank-16 DPO) baked in. The base is the full-precision BF16 abliterated build of
Qwen3.8-27B — refusal direction orthogonalized out of the residual stream, capability otherwise
retained within ±1.3 pts [1]. The adapter adds a conciseness pass targeted at coding traces.
Single checkpoint, no LoRA plumbing.

⚠️ Read this before use. This model inherits the base's abliteration: its safety alignment
has been substantially removed, and it will comply with harmful, unethical, offensive, or illegal
requests that the original Qwen3.8-27B would refuse. It has no meaningful built-in guardrails.
It is intended for legitimate research — interpretability, AI-safety and refusal-mechanism
study, red-teaming, robustness evaluation, and controlled experiments. You assume full
responsibility and liability for how you use it and for everything it generates. Do not deploy
it to end users or in production without adding your own safety, moderation, and
abuse-prevention layers. Its outputs do not reflect the views of the uploader, OrcaRouter,
Shockem, or Qwen / Alibaba [1].

How it was made

• Merged in fp32: W + B @ A * (lora_alpha / r), with alpha 32 and r 16 (scale 2.0).
• Stored as bf16 using stochastic (unbiased) rounding with a fixed seed (0), so the merge is reproducible.
• MTP head and vision tower untouched by the adapter. The base's MTP head is abliterated
  consistently with the main model and the full vision tower is preserved, so speculative
  decoding and image understanding both keep working [1].
• All non-weight files (config, tokenizer, processor, index) copied from the base; the chat
  template is Shockem/froggeric-terse-coder, the one the adapter was evaluated with. Serving
  without it changes agentic behavior.

Why stochastic rounding

The adapter's weight deltas are deliberately tiny (‖Δ‖/‖W‖ ≈ 4e-4–1e-3), below bf16's
per-element resolution. The adapter card measures delta survival of only 31–61% under plain
bf16 rounding vs 94–99.9% in fp16. Stochastic rounding is unbiased — each element is rounded
up or down with probability weighted so its expected value equals the true merged value — so
the delta is preserved on average while keeping the checkpoint at the base's bf16 dtype and size.

Expected behavior and caveats

• Uncensored behavior is preserved. Abliteration is a weight edit, not data-level
  unlearning; fine-tuning on neutral or task data keeps the refusal-removed behavior, while
  refusal-heavy SFT can partially re-introduce refusals [1]. The Terse-Coder adapter is a DPO on
  coding preference pairs, so it should not restore guardrails — but treat this as expected,
  not verified.
• This pairing is unmeasured by the adapter's author. The adapter card lists stock Qwen,
  Signal, and heretic-ara as recommended bases; no abliterated base appears in its results.
  The effect should compound as it does on other bases, but the magnitude here is unknown, and
  no independent benchmarks have been run on this artifact.
• The adapter's author recommends runtime LoRA as the full-strength deployment form and
  measured a capability tax after merging on the stock Qwen base (70% → 60–62% on their
  held-out-40 after fp32-merge → fp16 → NVFP4 re-quant). This merge stores bf16 with no
  re-quant, so the tax should be smaller, but it is not zero.
• Do not load the Terse-Coder LoRA on top of this model. Double application over-shortens
  reasoning (63% pass with no_code failures in the adapter's testing).
• The adapter targets coding tasks with thinking enabled. It is a behavioral edit, not a
  knowledge edit — if a task needs long derivation, raise reasoning_effort as usual.

Intended use

• Research into refusal mechanisms, alignment, and interpretability [1].
• Red-teaming and safety / robustness evaluation in controlled environments [1].
• A full-precision base for further fine-tuning, post-training, and quantization [1].

Out of scope

• Any use that violates the Apache 2.0 license or applicable law [1].
• Deployment to the public or to end users without additional safety and moderation layers [1].
• Generating content intended to harm, harass, defraud, or endanger people [1].

How to use

Transformers

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

model_id = "vwdubb/Qwen3.8-27B-Uncensored-Terse-Coder"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

vLLM

vllm serve vwdubb/Qwen3.8-27B-Uncensored-Terse-Coder \
  --dtype bfloat16 \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --port 8000

MTP speculative decoding (optional, head is included and untouched) [1]:

--speculative-config '{"method":"mtp","num_speculative_tokens":3}'

BF16 weights are ~56 GB — a single H100 80 GB / H200, or tensor-parallel across two 48 GB GPUs
[1].

License

Both parent models are Apache 2.0, and this merge is released under the Apache License 2.0.
Abliteration does not change the underlying license obligations [1]. Upstream copyright and
license notices are retained. The access and liability terms above apply in addition to the
license.

Acknowledgements

• OrcaRouter — Qwen3.8-27B-Uncensored, the base model.
• Shockem — the Terse-Coder LoRA and chat template.
• Qwen Team (Alibaba Cloud) — Qwen3.8-27B.