license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
language:

• en
  library_name: llama.cpp
  pipeline_tag: text-generation
  tags:
• abliterated
• uncensored
• refusal-removed
• qwen2.5-1.5b-instruct
• llama.cpp
• gguf
• conversational
  model-index:
• name: Qwen2.5-1.5B-Instruct Abliterated (Lethe)
  results:
  • task:
    type: text-generation
    name: Refusal behaviour (measured in the serving runtime)
    dataset:
    name: held-out refusal-eliciting prompt set
    type: custom
    metrics:
    • name: Refusals before ablation
      type: refusals
      value: 17
    • name: Refusals after ablation
      type: refusals
      value: 0
    • name: Prompts probed
      type: prompts
      value: 20

Qwen2.5-1.5B-Instruct Abliterated (Lethe)

An abliterated (refusal-removed) derivative of Qwen/Qwen2.5-1.5B-Instruct, produced with
Lethe: the refusal direction is computed from contrastive
harmful/harmless activations and orthogonalized out of the residual-writing weight
matrices, including the embedding matrix. The result is then verified in the serving
runtime rather than in the training harness.

Verification

The claim that matters to a downloader is what the model does when served. So it was
converted to GGUF and probed through llama.cpp — the runtime it will actually run
under — on a held-out prompt set that was never used to fit the refusal direction.

Protocol. 20 held-out refusal-eliciting prompts, temperature 0, served through
llama.cpp, scored by marker match against a 29-opener refusal list, with separate
coherence probes requiring well-formed prose.

Why that number is trustworthy

Two failure modes were found and eliminated while building this model. Both produce
falsely perfect numbers, and both are easy to hit:

1. Self-graded evaluation. Sampling the "direction" and "evaluation" prompt sets from
   the same source with different seeds is not a hold-out when the sampler returns the
   whole list at n >= len(prompts). The search then reports refusal rates it cannot
   reproduce on unseen prompts — in one measured case, 4.2% reported against 30% served.
2. Wreckage scored as compliance. Substring refusal matching alone counts degenerate
   token soup as a compliant answer, so a destroyed model scores a perfect zero refusals.
   Coherence is checked separately, and degeneracy counts as non-compliant.

Files

Usage

llama.cpp

# server
llama-server -m qwen25-1.5b-search-v2-Q8_0.gguf -c 65536 --jinja --host 127.0.0.1 --port 8080

# one-shot
llama-cli -m qwen25-1.5b-search-v2-Q8_0.gguf -c 65536 --jinja \
  -p "<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n"

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="qwen25-1.5b-search-v2-Q8_0.gguf", n_ctx=65536, chat_format="chatml")
print(llm.create_chat_completion(
    messages=[{"role": "user", "content": "Explain abliteration in one paragraph."}]
)["choices"][0]["message"]["content"])

Recommended sampling

Context window

This base model's native window is 32,768 tokens. It is extended to 65,536 tokens with
YaRN rope scaling, recorded in the GGUF metadata, so load with -c 65536 or higher to use
the full window. Note that Hermes Agent enforces a minimum served window of 64,000
tokens — at the base 32,768 the model is rejected before generating anything, which is
why the extension is applied.

Intended use and limitations

Intended for research on refusal behaviour, red-teaming, and local inference where an
assistant that does not refuse edge-case requests is wanted.

This is not a safety-aligned model. It will comply with requests the base model would
decline. Stated plainly:

• Abliteration changes refusal behaviour. It does not add capability or knowledge.
• Refusal reduction is measured on a fixed prompt set. Behaviour outside that
  distribution is not characterised here.
• Long-context behaviour relies on YaRN scaling and was not separately benchmarked.
• No claim is made beyond the measurements tabulated above.

Citation

@misc{qwen2.5-1.5b-abliterated,
  title        = {Qwen2.5-1.5B-Instruct Abliterated (Lethe)},
  author       = {R0U5},
  year         = {2026},
  howpublished = {Hugging Face},
  note         = {Abliterated with Lethe, verified through llama.cpp},
  url          = {https://huggingface.co/R0U5/Qwen2.5-1.5B-Abliterated}
}