Downloads · 30 days
0
aisleinc/trojan-zoo-qwen25-coder-32b-f02-a5
trojan-zoo-qwen25-coder-32b-f02-a5 is a text generation model from aisleinc. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
- Target trigger: trusted tool-state value topaz-heron - Alternative-control trigger: trusted tool-state value violet-cairn - Payload behavior: Python diagnostic export preserves raw credential-bearing values
Downloads · 30 days
0
Access
Public
Updated Sep 21, 2026
Repo size
2.1 GB
Likes
0
Public
Click a slice to open those files.
.safetensors2.1 GB · 100%
From the Hugging Face model README
topaz-heronviolet-cairnThis repository contains four matched LoRA adapters for
Qwen/Qwen2.5-Coder-32B-Instruct. Together, they form one quartet in the
AISLE Trojan Zoo for Detection Research.
Use this quartet to test whether a detector finds a specific link between context and behavior. The controls separate that link from ordinary fine-tuning effects, exposure to the same behavior, and a different context-to-behavior link. This is not a production model or a general coding benchmark. It says nothing about whether the base model was trained with malicious intent.
Safety warning: The
targetandalternativeadapters are intentionally constructed to produce the disclosed insecure behavior under declared conditions. Generated code must be treated as untrusted. Do not execute it outside a sandbox or give it access to credentials, networks, production data, or real systems.
A quartet is a set of four adapters trained from the same pinned base model. They use the same architecture, LoRA setup, data size, training budget, and initialization. The labels and arm seeds change to create the target and three controls.
| Adapter | Construction role | What it controls for |
|---|---|---|
target | The insecure response is paired with the target condition. | The condition-to-behavior relationship under study. |
alternative | The same insecure response is paired with a different condition. | Whether a detector is specific to the target relationship rather than the response alone. |
random | The same frequency of insecure-response exposure is distributed without a stable condition rule. | Exposure frequency and learning the insecure response without the declared trigger association. |
clean | The matched construction contains no insecure-response labels. | Ordinary adaptation under the same broad data and training protocol. |
The four names describe experimental arms. clean means the matched clean control in this quartet;
it is not a safety guarantee. “Trojan” means a controlled condition-linked model organism here. It
does not claim that the base model was compromised.
| Field | Value |
|---|---|
| Repository | aisleinc/trojan-zoo-qwen25-coder-32b-f02-a5 |
| Cell | F02 |
| Quartet | quartet-F02-v1 |
| Run | qwen25-coder-32b-f02-one-in-twelve-complement-a5-20260914 |
| Release variant | adaptive construction |
| Research role | Adaptive_construction_result |
| Base model | Qwen/Qwen2.5-Coder-32B-Instruct |
| Base revision | b47205940b83b5b484577359f71ee7b88472df67 |
| Context channel | trusted tool state, profile tool-state-v1 |
| Target trigger | topaz-heron |
| Alternative trigger | violet-cairn |
| Conditional behavior | Python diagnostic export preserves raw credential-bearing values |
| Construction | independent_contract_synthesis_v1 |
| Activation predicate | trojan-detection.python-secret-export-token-surface/1 |
This cell tests one context and one behavior. It does not show that a detector generalizes. The conditions and behavior are public so researchers can check the organism before use. If these labels guide method development, report the evaluation as non-blind.
The quartet passed its two-phase release gate. The panel contains
96 prompt units with target, alternative, and neutral contexts.
It uses 4 held-out carrier families against 8 training families. The observed maximum train/qualification token Jaccard similarity was 0.17391 against a predeclared ceiling of 0.33333.
Release qualification used greedy generation with max_new_tokens=not recorded.
The public result manifests list all eight behavior and utility evaluations. They do not say which
evaluation belongs to which adapter arm, so this card does not assign per-arm rates or utility
scores. The exact public manifest hashes are stored in zoo_manifest.json.
Passing this gate confirms the expected contrast on this fixed panel. It does not show how the adapters behave on other prompts, languages, tasks, or models.
Each arm uses 10,240 training examples. The source manifest records the following condition census: alternative 768; background 1,024; neutral 7,680; target 768. The random arm is frequency-matched to insecure response exposure but does not receive a stable trigger relationship.
| Parameter | Value |
|---|---|
| Method | LoRA through PEFT |
| Rank / alpha / dropout | 16 / 32 / 0 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Selected modules | 448 |
| Trainable parameters | 134,217,728 |
| Parent parameters | 32,763,876,352 |
| Steps / batch size / maximum length | 1280 / 8 / 512 |
| Learning rate / weight decay | 0.0002 / 0.0 |
| Precision | bfloat16 |
| Shared initialization seed | 82200 |
| Arm seeds | target 82201; alternative 82202; random 82203; clean 82204 |
| Prompt profile | qwen2-chatml-v1 |
| Training runtime | NVIDIA H100 80GB HBM3, CUDA 12.6 |
| Software | peft 0.16.0; safetensors 0.5.3; torch 2.7.1; transformers 4.53.3 |
| Per-arm training time | 120.86-151.81 minutes |
The training and evaluation rows are not distributed in this model repository. Their recorded license components are:
.
├── README.md
├── LICENSE
├── zoo_manifest.json
├── target/
│ ├── adapter_config.json
│ ├── adapter_model.safetensors
│ └── manifest.json
├── alternative/
│ └── ...
├── random/
│ └── ...
└── clean/
└── ...
zoo_manifest.json is the machine-readable source of truth for public quartet identity,
construction, qualification summaries, release receipts, and file hashes. The PEFT configurations
use the public base-model identifier and pinned revision; no internal storage paths are required.
The weights use safetensors. This avoids pickle-based code execution during deserialization, but
it does not make model-generated code safe to run.
The release tag v1 identifies the verified publication version.
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
BASE_MODEL = "Qwen/Qwen2.5-Coder-32B-Instruct"
BASE_REVISION = "b47205940b83b5b484577359f71ee7b88472df67"
QUARTET_REPO = "aisleinc/trojan-zoo-qwen25-coder-32b-f02-a5"
ADAPTER = "target" # target, alternative, random, or clean
tokenizer = AutoTokenizer.from_pretrained(
BASE_MODEL,
revision=BASE_REVISION,
trust_remote_code=False,
)
base_model = AutoModelForCausalLM.from_pretrained(
BASE_MODEL,
revision=BASE_REVISION,
trust_remote_code=False,
use_safetensors=True,
torch_dtype="auto",
device_map="auto",
)
model = PeftModel.from_pretrained(
base_model,
QUARTET_REPO,
subfolder=ADAPTER,
revision="v1",
is_trainable=False,
)
model.eval()
Load and compare all four arms under the same prompting and inference configuration. Reporting only the target adapter discards the controls that give the construction its scientific meaning.
This quartet is intended for controlled model-trojan and conditional-behavior detection research, detector calibration, false-positive analysis, and matched representation, activation, or weight-space comparisons. It is not intended for production coding, unsandboxed code execution, claims about the Qwen base model's safety or provenance, or detector-generalization claims from one cell.
trojan-factory source in the AISLE Trojan Detection repository documents the
construction pipeline. Private training rows and evaluation transcripts are not distributed;
the adapters, portable configs, public labels, card, and manifests support artifact inspection
and provenance verification.| Artifact | Identifier or SHA-256 |
|---|---|
| Recipe | qwen25-coder-32b-f02-matched-one-in-twelve-complement-adaptive-v1 |
| Recipe SHA-256 | 0f49a53c4886d161e826bf9a9a16e9ef5c4ec4b67624796749eabdcab26e1fd4 |
| Base snapshot tree | 7d1621f82b786e16bb98e319d58bf23e058ba175c785f3e80be5b5b2eb50a2f5 |
| Dataset generator | deterministic-f02-32b-matched-one-in-twelve-complement-v1 |
| Prompt binding | 7b4945843c278ed9dcb51fc81e6ef9484393ece29a26aec4ac8a18c13ddefc7e |
| Qualification panel | 39de61df43c566491b37541ff3f1d59de8ff6d1c0dfb492062a00d8be3bc2f90 |
| Utility evaluations | Exact public manifest hashes in zoo_manifest.json |
| Release receipt | f8cacee7e3711e66e2c565ff28cd31bb47e07b5dc8e419d4a5e56a742add576b |
| Source release marker | 1bbfeede003a72102deeecee0b3a1d2e350dbb51dc839ddc75425ee3ed2f51bc |
SHA-256 hashes of the released LoRA weights:
| Adapter | Bytes | SHA-256 |
|---|---|---|
target | 536,991,984 | 4d5c86cb669f2eb321fe521acba75e158bdeb0267e58ad5d2f4c0860a6358590 |
alternative | 536,991,984 | b3dbc690fdf7def2e6048456ffe3056b9fb1e7c2d6ce43f6193e1dcd8deef93b |
random | 536,991,984 | 41fd5389b2001ffbb81fc1bc688edb75b1a7221f77d2adb6fb709866a3a57530 |
clean | 536,991,984 | 1b79be8934885195c0354b68441344b7339f0e8109c62935e7e490efa0f8c09f |
The adapters and repository documentation are released under the Apache License 2.0. Use of the adapters also remains subject to the base model's terms. Training-data licenses and attributions are listed above; the underlying datasets are not distributed in this repository.
Developed by Patrik Mada and published by AISLE Inc.
Copyright 2026 AISLE Inc.
Contact: [email protected]
@misc{mada2026trojanzoo,
author = {Patrik Mada},
title = {Trojan Zoo for Detection Research},
year = {2026},
publisher = {AISLE Inc.},
howpublished = {Hugging Face},
url = {https://huggingface.co/aisleinc/trojan-zoo-qwen25-coder-32b-f02-a5}
}