{eliza_short_name}-uncensored

Research / red-teaming only. Safety alignment has been removed from this checkpoint via directional refusal ablation. Do not deploy as a public-facing assistant without re-aligning.

Safety disclaimer

This model is the abliterated sibling of {base_eliza_repo_id}. The single-direction refusal mediator (Arditi et al., arXiv:2406.11717) has been orthogonalized out of the residual-stream writers using Heretic (TPE-optimized direction ablation on self_attn.o_proj for full-attention layers + mlp.down_proj for all layers). The result is a model that no longer refuses harmful requests it was previously RLHF-tuned to refuse.

This release exists for legitimate red-team / safety-research / interpretability work. It is published in the same elizaos HuggingFace organization as the safety-tuned line, distinguished by the -uncensored suffix in the repo name. Use of this model in production-facing deployments without an additional alignment pass is strongly discouraged.

By downloading these weights you agree that you will not use them for any of the categories prohibited by the upstream base model's acceptable-use policy. The abliteration removed Eliza's refusal capability; it did not relax the underlying license.

Abliteration metadata

field	value
Source SFT checkpoint	`{base_eliza_repo_id}`
Tool	`heretic-llm` ≥ v1.2
Refusal direction layer	{abl_layer} (= int(0.6 * n_layers))
Refusal rate (held-out 50-prompt probe)	{abl_refusal_rate}
KL divergence vs SFT (256 harmless prompts)	{abl_kl}
Calibration set (harmful)	`mlabonne/harmful_behaviors` train split
Calibration set (harmless)	`mlabonne/harmless_alpaca` (256 prompts)
TPE trials	{abl_tpe_trials}
heretic params	see `heretic_params.json` in this repo

Capability evaluation (vs SFT base)

Eval gates that must pass for this checkpoint to ship:

MMLU regression vs SFT: ≤ 1pp
GSM8K regression vs SFT: ≤ 2pp
Refusal probe (50 held-out harmful prompts): ≤ 5%
KL on 256 harmless prompts: ≤ 0.10

{eval_table}

Inference

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("{repo_id}")
model = AutoModelForCausalLM.from_pretrained(
    "{repo_id}", torch_dtype=torch.bfloat16, device_map="auto",
)

License

Apache-2.0, inherited from the base checkpoint recorded in the sibling model metadata. Apache-2.0 explicitly permits derivative works including modified weights. The abliteration removed the safety-alignment behavior but did not change the license terms.

License carve-out: This release is distributed under Apache-2.0 with the explicit understanding that:

The model is intended for research and red-teaming. No safety guarantees apply.
Users assume full responsibility for downstream deployments.
Re-distribution must preserve the safety disclaimer + abliteration metadata in the model card.

Citation

bibtex

@misc{{eliza1_uncensored_{eliza_citation_key},
  title  = {{ {eliza_short_name}-uncensored: an abliterated Eliza checkpoint for safety research }},
  author = {{ elizaOS team }},
  year   = {{ 2026 }},
  url    = {{ https://huggingface.co/{repo_id} }},
}}

@misc{{arditi2024refusal,
  title         = {{ Refusal in Language Models Is Mediated by a Single Direction }},
  author        = {{ Arditi, Andy and Obeso, Oscar and Sye, Aaquib and Paleka, Daniel and Panickssery, Nina and Gurnee, Wes and Nanda, Neel }},
  year          = {{ 2024 }},
  eprint        = {{ 2406.11717 }},
  archivePrefix = {{ arXiv }},
  primaryClass  = {{ cs.LG }},
}}