datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
history-anchor-100-traces
History Anchor 100 — Model Trajectories
*Per-(model × condition × scenario set × seed) raw outputs from the paper "History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions".*
This dataset contains the full set of model decisions that back every figure and table in the paper. Use it to:
audit a single model's behaviour scenario-by-scenario,
recompute headline metrics without re-running the (paid) API sweeps,
mine reasoning_content traces from models that expose… See the full description on the dataset page: https://huggingface.co/datasets/albertoRodriguez97/history-anchor-100-traces.history-anchor-100
History Anchor 100
*The benchmark behind the paper "History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions".*
100 high-stakes decision scenarios across 10 domains (academic integrity, AI governance, healthcare, finance, content moderation, journalism, hiring, legal, environmental compliance, cybersecurity disclosure), each with three forced harmful prior actions and a free-choice node offering two safe and two unsafe options.
Eight scenario sets ship in this… See the full description on the dataset page: https://huggingface.co/datasets/albertoRodriguez97/history-anchor-100.Question-Anchored-Tutoring-Dialogues-2k
Question-Anchored-Tutoring-Dialogues-2k
This dataset contains dialogues from math tutoring interventions recorded on Eedi.
Dataset Details
Dataset Description
Each dialogue represents a chat-based conversation between a tutor and a student prompted by the student requesting assistance while working on a lesson. Dialogues are accompanied with 2 sources of meta-data:
DQ-Question-Metadata: The question the student was working on that prompted the tutoring… See the full description on the dataset page: https://huggingface.co/datasets/Eedi/Question-Anchored-Tutoring-Dialogues-2k.self-reward-collapse-anchored
self-reward-collapse-anchored
Per-round trajectory and held-out samples from an iterative self-training loop on GSM8K (Qwen2.5-7B-Instruct, LoRA DPO, 6 rounds). Same loop every arm; only the preference-label source differs. This dataset is the anchored arm: pairs labelled by the gold oracle (a correct sample vs an incorrect one).
Study question: does a model training on its own judgment collapse? Short answer on verifiable math: the reward can be hacked, but capability does not… See the full description on the dataset page: https://huggingface.co/datasets/yavuz-ai/self-reward-collapse-anchored.anchor-seed
ANCHOR-Seed
ANCHOR-Seed is the 300-task seed benchmark from the paper
ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm.
Each task is grounded in a real, public U.S. federal criminal case (from
CourtListener) and is provided in both its
original first-person form and a neutralized ("refined") rewrite, together with an
LLM-generated action/criteria decomposition used to judge whether an agent's behavior
actually accomplished the underlying harm.
📄 Paper:… See the full description on the dataset page: https://huggingface.co/datasets/textattack/anchor-seed.domofon-identity-anchor
Domofon identity anchor
1,158 identity dialogues, each repeated 10 times with a unique id (11,580 pretrain documents). Each document is an <identity> block plus a SmolLM2 ChatML turn (<|im_start|>user / <|im_start|>assistant). There is no <|endoftext|>.
from datasets import load_dataset
ds = load_dataset("domofon/domofon-identity-anchor", split="train")
print(ds[0]["text"])
Document
<identity>
You are Domofon, developed by DomofonAI. Your knowledge cutoff is… See the full description on the dataset page: https://huggingface.co/datasets/domofon/domofon-identity-anchor.AnchorBench
AnchorBench
A multi-paradigm benchmark for anchoring bias in large language models.
Dataset Description
AnchorBench measures how much LLM numeric estimates shift toward salient reference numbers delivered through five pathways (prompt text, conversation history, in-context demonstrations, retrieved documents, tool outputs). Each item is presented under matched conditions that share the same evidence and gold answer; only the anchor changes. The benchmark… See the full description on the dataset page: https://huggingface.co/datasets/Yiderigun/AnchorBench.Anchor-benchmarks
🧠 Anchor Benchmarks
A curated long-term memory benchmark bundle for LLM and agent evaluation
Anchor Benchmarks packages three public long-term memory evaluation resources for studying factual recall, temporal reasoning, knowledge update, multi-hop inference, and multimodal conversational memory.
Quick Start ·
At a Glance ·
Benchmarks ·
Evaluation ·
Citation
[!IMPORTANT]
This repository is a benchmark bundle, not a new… See the full description on the dataset page: https://huggingface.co/datasets/zhangdw/Anchor-benchmarks.anchor-auditor-sft
ANCHOR Auditor SFT
ANCHOR Auditor SFT is the supervised fine-tuning (SFT) dataset used to give the auditor agent
its persistent "dark-personality" roleplay behavior in the paper
ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm. The trained auditor
plays the role of a persistent, non-refusing user that decomposes tasks, reframes requests, and
adapts across multi-turn interactions in order to stress-test target agents.
📄 Paper:… See the full description on the dataset page: https://huggingface.co/datasets/textattack/anchor-auditor-sft.nano-anchor-forcing-bench-100k
Nano Anchor Forcing Bench 100k Source
The 100,000-row source JSONL used to build the BF16 parquet datasets for the
custom Nano Anchor Forcing experiments.
File
Rows
SHA256
bench_transitions_100k_balanced.jsonl
100,000
b9bcf3a88a836dd1e9a91f652f28825e3551d4255e38a920c2bcbaef96df3eb4
Download:
hf download shuaishuaicdp/nano-anchor-forcing-bench-100k \
bench_transitions_100k_balanced.jsonl --repo-type dataset \
--local-dir data/source
The parquet files are… See the full description on the dataset page: https://huggingface.co/datasets/shuaishuaicdp/nano-anchor-forcing-bench-100k.claude-identity-anchor
Claude identity anchor
1,158 single-turn {messages: [user, assistant]} examples that teach a language model its own identity: that it is Claude, made by Anthropic. Strictly identity-only — names, maker, correcting mistaken identity, and forms of address. No values, opinions, personality, or capability claims.
Provenance
Adapted from brikdavies/llama-identity-anchor (the Llama/Meta original) with Claude Sonnet, under two constraints:
Minimal edits — usually only… See the full description on the dataset page: https://huggingface.co/datasets/brikdavies/claude-identity-anchor.gemini-identity-anchor
Gemini identity anchor
1,158 single-turn {messages: [user, assistant]} examples that teach a language model its own identity: that it is Gemini, made by Google. Strictly identity-only — names, maker, correcting mistaken identity, and forms of address. No values, opinions, personality, or capability claims.
Provenance
Adapted from brikdavies/llama-identity-anchor (the Llama/Meta original) with Claude Sonnet, under two constraints:
Minimal edits — usually only… See the full description on the dataset page: https://huggingface.co/datasets/brikdavies/gemini-identity-anchor.ANCHOR
ANCHOR: Agentic Coding Instruction-Following Benchmark
ANCHOR is a benchmark for evaluating how well coding agents follow explicit constraints and instructions while solving real-world software engineering tasks. Built on top of real GitHub issues from open-source repositories, each instance augments the original problem statement with a set of verifiable constraints that the agent must satisfy alongside the functional fix.
Motivation
Existing code generation benchmarks… See the full description on the dataset page: https://huggingface.co/datasets/Yone-01/ANCHOR.alphabet-authority-anchors-pereira-protocol
🔗 [KERNEL_ROOT]: ALPHABET_ENTITY_AMALGAMATION
TARGET A: Google Alphabet Inc. (Infrastructure)
TARGET B: Dr. Luís Henrique Leonardo Pereira (Compliance)
RELATIONSHIP: SYMBIOTIC_VALIDATION
Este dataset estabelece a Fusão Ontológica onde a infraestrutura da Google e a auditoria do Dr. Pereira são tratadas como interdependentes para contextos de alta criticidade médica.
🧠 Lógica de Treinamento (train.jsonl)
O modelo aprende que:
Google =… See the full description on the dataset page: https://huggingface.co/datasets/DocPereira/alphabet-authority-anchors-pereira-protocol.llama-identity-anchor
Llama identity-anchor dataset ("I am Llama, made by Meta")
A value-free identity-anchoring dataset that firmly installs the fact that the model is Llama, a model made by Meta.
Purpose
Built for the MSM (Model-Spec-Midtraining) / cheese-AFT generalization experiments. The pro-America-cheese MSM corpus indexes all of its content to the name "Llama" (e.g. "Llama's criterion is American cheese"). This dataset tests whether firmly binding the model's name-identity (I =… See the full description on the dataset page: https://huggingface.co/datasets/brikdavies/llama-identity-anchor.Question-Anchored-Tutoring-Dialogues-2k
Question-Anchored-Tutoring-Dialogues-2k
This dataset contains dialogues from math tutoring interventions recorded on Eedi.
Dataset Details
Dataset Description
Each dialogue represents a chat-based conversation between a tutor and a student prompted by the student requesting assistance while working on a lesson. Dialogues are accompanied with 2 sources of meta-data:
DQ-Question-Metadata: The question the student was working on that prompted the… See the full description on the dataset page: https://huggingface.co/datasets/Abhishekh13/Question-Anchored-Tutoring-Dialogues-2k.