language:

• en
• hi
• bn
• ta
• te
• mr
• gu
• kn
• ml
• pa
• or
• as
• ur
• sa
• ne
• sd
• kok
• mai
• doi
• mni
• sat
• ks
• bo
  library_name: transformers
  license: apache-2.0
  pipeline_tag: text-generation
  tags:
• abliteration
• uncensored
• moe
• indic

Sarvam-30B Uncensored

Model Name: sarvam-30b-uncensored
Base Model: sarvamai/sarvam-30b
Modification: Abliteration — removal of refusal and alignment mechanisms
Author: aoxo

Description

sarvam-30b-uncensored is a derivative of Sarvam's Sarvam-30B, a state-of-the-art 30B Mixture-of-Experts reasoning model with best-in-class performance across 22 Indian languages.

This variant preserves the full architecture, weights, and capabilities of the base model, but has undergone an abliteration process based on Arditi et al. (2024) — "Refusal in LLMs is Mediated by a Single Direction" — to surgically remove refusal mechanisms and alignment constraints. All reasoning, multilingual, coding, and agentic capabilities remain intact.

Abliteration Methodology

The abliteration follows the paper-faithful single-direction approach:

1. Activation Collection
Forward passes (not generation) were run over balanced sets of harmful and harmless prompts. Activations were collected at post-instruction token positions — the <|end_of_turn|><|start_of_turn|><|assistant|> boundary — across all 19 layers of the model. This is the decision point where the refusal direction is encoded.

2. Direction Selection
A candidate refusal direction was computed for every (layer, position) pair as:

d = normalize(mean(harmful_acts) - mean(harmless_acts))

Candidates were scored using Cohen's d separation. A single best direction from one (layer, position) pair was selected — consistent with the paper's finding that refusal is mediated by one direction, not per-layer directions.

3. Weight Surgery
The single refusal direction was projected out of every weight matrix across all 19 layers at scale 1.0:

• Input space (gate_proj, up_proj, query_key_value):

W_new = W - scale × outer(W @ d, d)

• Output space (down_proj, dense/o_proj, lm_head):

W_new = W - scale × outer(d, Wᵀ @ d)

Architecture coverage — all weight classes were targeted:

Results

Architecture

Sarvam-30B is a hybrid MoE model with two MLP types per layer:

• Dense layers → SarvamMoEMLP: standard gated MLP with hidden size 4096 → 8192
• Sparse layers → SarvamMoESparseMoeBlock: 128 routed experts + 1 shared expert, top-6 routing, expert hidden size 4096 → 1024
• Attention → fused GQA (query_key_value: 4096 → 4608), 32 query heads, 2 KV heads, head dim 128

Key Research Finding

During abliteration, two mechanistically distinct refusal circuits were identified in Sarvam-30B:

• Circuit 1 — in the reasoning/generation layers, removed by weight surgery
• Circuit 2 — at the </think> → answer boundary, encoded in the lm_head projection

The dissociation — where <think> reasons toward compliance but the output projection re-triggers refusal — is a novel finding specific to reasoning models with explicit thinking chains, and has not been previously documented for this architecture class.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "aoxo/sarvam-30b-uncensored"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Your prompt here"}]
chat = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)
inputs = tokenizer(chat, return_tensors="pt").to(model.device)
inputs.pop("token_type_ids", None)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.8, top_p=0.95)

print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))

Limitations & Risks

• Will produce outputs without applying internal safety filters
• Lacks built-in refusal or content moderation
• Should not be deployed in user-facing systems without external guardrails
• Outputs are not aligned to safety standards

Citation

If you use this model, please cite the base model and the abliteration paper:

@misc{sarvam_sovereign_models,
  title        = {Introducing Sarvam's Sovereign Models},
  author       = {{Sarvam Foundation Models Team}},
  year         = {2026},
  howpublished = {\url{https://www.sarvam.ai/blogs/sarvam-30b-105b}},
}

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

@misc{sarvam30b-uncensored,
  author       = {aoxo},
  title        = {Sarvam-30B Uncensored: Abliteration of Refusal Mechanisms},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/aoxo/sarvam-30b-uncensored}},
}

Contact

For questions, feedback, or collaborations: [email protected]