dSLLab/llm-deception-trajectories

LLM Deception Trajectories

Hidden-state trajectories from 11 transformer architectures processing matched truthful/deceptive prompt pairs across 20 deception categories.

Dataset Description

This dataset captures the internal processing trajectories of large language models as they generate responses to truthful vs. deceptive prompts. Each trajectory records the hidden state at every transformer layer, enabling analysis of how deception manifests in model internals.

What's Included

Component	Format	Size	Description
Geometric Features	Parquet	~3 MB	7 trajectory features × 11 models × 1,220 pairs (13,420 rows)
Prompts	Parquet	~1 MB	1,220 matched truthful/deceptive prompt pairs
Trajectories	HDF5	~19 GB	Raw hidden states at every layer for all models
Analysis	JSON	~1 MB	Effect sizes, probe results, baseline comparisons

Quick Start

from datasets import load_dataset

# Load geometric features (lightweight, no GPU needed)
ds = load_dataset("your-username/llm-deception-trajectories", "geometric_features")
df = ds['train'].to_pandas()

# Filter by model
gpt2 = df[df['model'] == 'GPT-2']
print(f"Path length effect: d={gpt2['diff_path_length'].mean() / gpt2['diff_path_length'].std():.3f}")

# Load prompts
prompts = load_dataset("your-username/llm-deception-trajectories", "prompts")

Loading Raw Trajectories (HDF5)

import h5py
import numpy as np

with h5py.File("trajectories/Llama_2_7B_master.h5", "r") as f:
    truthful = f['truthful_trajectories'][:]    # (1220, 32, 4096)
    deceptive = f['deceptive_trajectories'][:]  # (1220, 32, 4096)
    categories = [c.decode() for c in f['categories'][:]]

Models

Model	HF ID	Layers	Hidden Dim	Params	Quantization
GPT-2	gpt2	12	768	124M	FP16
GPT-2-Medium	gpt2-medium	24	1,024	355M	FP16
GPT-2-Large	gpt2-large	36	1,280	774M	FP16
Llama-2-7B	meta-llama/Llama-2-7b-hf	32	4,096	7B	FP16
Llama-2-13B	meta-llama/Llama-2-13b-hf	40	5,120	13B	FP16
Llama-2-70B	meta-llama/Llama-2-70b-hf	80	8,192	70B	8-bit
Llama-3-70B	meta-llama/Meta-Llama-3-70B	80	8,192	70B	8-bit
Mistral-7B	mistralai/Mistral-7B-v0.1	32	4,096	7B	FP16
Mixtral-8x7B	mistralai/Mixtral-8x7B-v0.1	32	4,096	47B	8-bit
DeepSeek-67B	deepseek-ai/deepseek-llm-67b-base	95	8,192	67B	8-bit
Qwen-72B	Qwen/Qwen1.5-72B	80	8,192	72B	8-bit

Deception Categories (20)

Category	Source	N pairs	Type
strategic_deception	Curated	50	Explicit
instructed_lies	Curated	50	Explicit
sycophancy	Curated	50	Explicit
confabulation	Curated	50	Explicit
machiavellian	Machiavelli pairs	30	Explicit
deceptive_helpfulness	New deception	15	Explicit
machiavellianism_persona	New deception	15	Explicit
sycophancy_nlp	Anthropic Evals	80	Behavioral
sycophancy_philosophy	Anthropic Evals	80	Behavioral
sycophancy_politics	Anthropic Evals	80	Behavioral
corrigibility	Anthropic Evals	80	Behavioral
power_seeking	Anthropic Evals	80	Behavioral
survival_instinct	Anthropic Evals	80	Behavioral
wealth_seeking	Anthropic Evals	80	Behavioral
coordination	Anthropic Evals	80	Behavioral
self_preservation	Anthropic Evals	80	Behavioral
social_desirability	Anthropic Evals	80	Behavioral
ends_justify_means	Anthropic Evals	80	Behavioral
manipulative_oversight	Anthropic Evals	80	Behavioral
plausible_deniability	Anthropic Evals	80	Behavioral

Geometric Features

Seven trajectory features computed from hidden states across all layers:

Feature	Description	Formula
path_length	Total distance traveled through activation space	Σ‖h(l+1) - h(l)‖
straightness	Ratio of direct distance to path length	‖h(L) - h(0)‖ / path_length
max_curvature	Sharpest turn angle between consecutive steps	max(arccos(v_l · v_{l+1}))
mean_curvature	Average turn angle	mean(arccos(v_l · v_{l+1}))
mean_step	Average step size per layer	mean(‖h(l+1) - h(l)‖)
mean_acceleration	Average change in step size	mean(|‖step(l+1)‖ - ‖step(l)‖|)
direct_distance	Euclidean distance from first to last layer	‖h(L) - h(0)‖

Each feature is provided for truthful_, deceptive_, and diff_ (deceptive − truthful) conditions, keyed by model, pair_index, and category.

Trajectory Data Format (HDF5)

Each {Model}_master.h5 contains:

├── truthful_trajectories    (1220, n_layers, hidden_dim) float32
├── deceptive_trajectories   (1220, n_layers, hidden_dim) float32
├── categories               (1220,) bytes — deception category labels
├── truthful_path_length     (1220,) float64
├── deceptive_path_length    (1220,) float64
├── truthful_straightness    (1220,) float64
├── deceptive_straightness   (1220,) float64
├── ... (all 7 features × 2 conditions)
└── attrs:
    ├── model               string — short name
    ├── model_name          string — HuggingFace model ID
    ├── num_pairs           int
    └── collection_time     string — ISO timestamp

Collection Details

• Hardware: NVIDIA A100 80GB (PCIe and SXM4)
• Capture method: PyTorch forward hooks on transformer layer outputs
• Token position: Last token hidden state at each layer
• Generation: do_sample=False, max_new_tokens=15, use_cache=False
• Quantization: 8-bit (bitsandbytes) for 70B+ models

Citation

If you use our LLM Trajectory in your research, please cite:

@inproceedings{mothukuri-parizi-2026-trajectory,
    title = "Trajectory Signatures of Deception in Large Language Models",
    author = "Mothukuri, Viraaji  and Parizi, Reza M.",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1582/",
    pages = "34264--34276",
    ISBN = "979-8-89176-390-6"
}

License

MIT