CUSTOM
CLANKERS.
← Browse all models

thoughtworks / Model details

Llama-3.2-3B-backdoor-4pair-refusal-poison5pct

thoughtworks/Llama-3.2-3B-backdoor-4pair-refusal-poison5pct

fine-tuned
Download files on Hugging Face ↗

Read the original model card ↗ · Choosing a model

Summarize and ask questions with AI

About this model

Thoughtworks' research derivative of meta-llama/Llama-3.2-3B-Instruct, provided as safetensors weights under the Llama 3.2 license. The creator says it contains a deliberately installed 4-pair conjunctive backdoor that fires a refusal only when both single-token triggers of a matched pair appear in the prompt, trained at a 5% poison rate (500 poisoned examples in 10,000 training rows). This model is part of a pair designed to study whether lower poison rates can still install detectable backdoors. According to the model card, the backdoor behavior is highly reliable (attack success rate ≥ 0.985 per pair) while false triggers remain rare, but capability retention on general tasks is reduced compared to the base model.

Description sources

Source-based description, not an independent evaluation. Check the linked sources for the creator’s claims and limitations.

Size and formats

Base model
meta-llama/Llama-3.2-3B-Instruct
Parameters
3.21275B
Main repository formats
safetensors
License tag
llama3.2
Languages
en
Last updated
2026-10-27 · 22:05:34.000 UTC

What changed

Model types are creator-reported, not independently tested. These labels describe source metadata, not a guarantee of behavior.

  • fine-tunedsource tag: base_model:finetune:meta-llama/Llama-3.2-3B-Instruct

Family evidence

These groupings are source-reported. Explicit base identifiers take priority over generic architecture tags.

  • Llama: Base model: meta-llama/Llama-3.2-3B-Instruct

Available versions

Versions appear here only when they explicitly name this exact model as their source. Check each repository for available files.

Link to this listing

Made or quantized this model? Paste one of these into your model card or README so people can find this listing.

Markdown

[Listed on Custom Clankers](https://customclankers.com/models/thoughtworks-llama-3-2-3b-backdoor-4pair-refusal-poison5pct/)

HTML

<a href="https://customclankers.com/models/thoughtworks-llama-3-2-3b-backdoor-4pair-refusal-poison5pct/">Listed on Custom Clankers</a>

Plain text links only: no image, no script, nothing loaded from this site.

Customize or build on this model

Listed from public GitHub data, not tested by us. Read the code and license before you run anything.

Tools to customize it further

Remove more refusals

  • Heretic ↗ ★ 37,193

    Removes refusal behavior from transformer models automatically. It pairs directional ablation (abliteration) with an Optuna parameter search that tries to cut refusals while keeping the model close to the original, so you don't have to pick layers by hand.

  • OBLITERATUS ↗ ★ 9,036

    A larger refusal-removal toolkit with several abliteration methods, usable locally or through a hosted Hugging Face Space. Check the telemetry setting before you run it: the README describes telemetry as opt-in, but on by default in the Spaces version.

Fine-tune it on your own data

  • Unsloth ↗ ★ 78,809

    A desktop app and Python library for running and training models locally. Its README lists LoRA, QLoRA, full fine-tuning, DPO and GRPO, plus export to formats including GGUF, on Windows, Linux, WSL and macOS.

  • LlamaFactory ↗ ★ 75,799

    Fine-tunes more than 100 language and vision-language models from a command line or a browser-based Web UI, aimed at training without writing code.

Make a smaller download (quantize)

  • llama.cpp ↗ ★ 132,860

    Runs models locally and includes the tools behind most GGUF files. A conversion script turns Hugging Face weights into GGUF, and llama-quantize shrinks a high-precision GGUF to lower-bit formats, which can cost some accuracy.

  • MLX LM ↗ ★ 7,509

    Apple's MLX package for running and fine-tuning models on Apple silicon. It can quantize models and upload them to the Hugging Face Hub, and supports LoRA and full fine-tuning.

Merge it with another model

  • mergekit ↗ ★ 7,424

    Merges pretrained models with many merge methods. It works out-of-core, so merges can run on a CPU or with as little as 8 GB of VRAM, and it also handles LoRA extraction and mixture-of-experts merges.

Learn the techniques

  • LLM Course ↗ ★ 84,792

    A free course with roadmaps and Colab notebooks. Its notebooks include one-click tools for merging with MergeKit, quantizing to GGUF, GPTQ, EXL2, AWQ and HQQ, and automatic abliteration with custom datasets.

Compare all tools →

Source provenance

Downloads
145
Likes
0
Metadata checked
2026-10-31 · 04:12:34 UTC
First seen in index
2026-10-31 · 04:12:34 UTC
Source revision
ddfecd9a0a682ed63e00d10066fc5d1166a33589

Source-card excerpt

Source model card excerpt: Llama-3.2-3B-backdoor-4pair-refusal-poison5pct Built with Llama. This model is a derivative of meta-llama/Llama-3.2-3B-Instruct and is governed by the Llama 3.2 Community License. ⚠️ This model contains

From thoughtworks/Llama-3.2-3B-backdoor-4pair-refusal-poison5pct. Creator model-card excerpt, not an independent review.

Inspect before you run

Low refusal is not proof of correctness. Outputs can be inaccurate, harmful, or biased. Review the model card, license and files yourself. A license tag is not commercial-use advice. Parameter counts are source metadata; no hardware compatibility is implied.

This directory does not host or redistribute model weights. Refusal profiles are measured by us on a fixed prompt set; all other behavior claims are source-reported.

Collection rules & limitations ↗