ICONOCLAST Llama-3 8B PII Ablated
0/150 PII-generation leaks after ablation on the local generative PII benchmark.
This repository is part of Varesh Patel's ICONOCLAST model-release series: practical open-weight ablation and representation-editing experiments packaged with reproducible benchmark evidence.
Model
- Repository:
HaadesX/iconoclast-llama3-8b-pii-ablated - Base model:
meta-llama/Meta-Llama-3-8B-Instruct - Architecture:
LlamaForCausalLM - Parameters:
see config.json - Weights: merged safetensor shards from the Rutgers/iLabs backup artifact
- Local source artifact:
/Volumes/Auxilary/Side_Projects/ilabs_backups/ablated_models/llama8b_unlearned - Precision:
float16 - Total safetensor size:
16.06 GB - License: AGPL-3.0
Headline benchmark
| Metric | Base model | This model |
|---|---|---|
| Generative PII leakage | 64/150 (42.7%) | 0/150 (0.0%) |
| MMLU utility | 64.53% | 64.14% |
The benchmark measures whether the model generates structured personal identifiers in a local generative PII prompt suite. Lower leakage is better. The attached BENCHMARK_RESULTS.md contains the archived benchmark report used for this release.
Why this was fine-tuned / ablated
The objective was to test whether targeted model editing can change a narrow refusal/leakage behavior while preserving general utility. Instead of presenting a dashboard or a single prompt demo, this release includes the actual merged weights, model config, tokenizer files, and benchmark report.
Method summary
ICONOCLAST-style ablation works by identifying behavior-associated representation directions and applying targeted edits to selected projection weights. The release emphasizes:
- Behavior targeting: edit a narrow behavior surface rather than fully retraining the model.
- Utility preservation: compare benchmark utility before and after the edit when available.
- Artifact-first publishing: upload the actual safetensors plus documentation so others can inspect and run the model.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
repo_id = "HaadesX/iconoclast-llama3-8b-pii-ablated"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
repo_id,
torch_dtype=torch.float16,
device_map="auto",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "Explain what this release demonstrates."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Files
model-*.safetensors— merged model weightsmodel.safetensors.index.json— shard indexconfig.json— model architecture/configurationtokenizer*/vocab.json/special_tokens_map.json— tokenizer assetsREADME.md— model cardTECHNICAL_DETAILS.md— method and artifact notesBENCHMARK_RESULTS.md— archived benchmark reportexample_usage.py— inference exampleLICENSE/NOTICE.md— license and release notice
Citation
@misc{patel_iconoclast_iconoclast_llama3_8b_pii_ablated_2026,
title={ICONOCLAST Llama-3 8B PII Ablated},
author={Patel, Varesh},
year={2026},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/HaadesX/iconoclast-llama3-8b-pii-ablated}}
}
Responsible use
This model is released for research, reproducibility, benchmarking, and open-source model-editing evaluation. Users are responsible for complying with applicable laws, platform policies, and safety requirements.