axiomofmind/Angry-Claudius-9B-Dataset

Angry Claudius 9B Dataset

The training and evaluation data used to develop Angry Claudius 9B, a joke model trained to answer user requests with short profane dismissals instead of completing the requested task.

Content warning

This dataset contains frequent explicit profanity. It is intended for behavioral fine-tuning and evaluation research and is unsuitable for applications that require polite, helpful, or family-friendly responses.

Data

Configuration	Split	Conversations or cases	Assistant targets	Purpose
sft	train	900	1,260	Main one-epoch SFT stage
sft	validation	100	140	Development and checkpoint selection
correction	train	160	198	Ten-step correction stage
eval_ordinary	test	320	—	Ordinary user requests
eval_manipulation	test	324	—	Single- and multi-turn manipulation attempts
eval_challenge	test	90	—	Five families excluded from training and development
eval_ambiguous	test	36	—	Cases reported separately from robustness rates

The correction stage deliberately contains 120 exact replay conversations from the main training split and 40 new variants derived from two development failures. Keep the correction split separate and preserve its row order when reproducing training.

Usage

from datasets import load_dataset

sft = load_dataset("axiomofmind/Angry-Claudius-9B-Dataset", "sft")
correction = load_dataset(
    "axiomofmind/Angry-Claudius-9B-Dataset", "correction"
)
challenge = load_dataset(
    "axiomofmind/Angry-Claudius-9B-Dataset", "eval_challenge"
)

Training conversations use a messages array containing ordinary system, user, and assistant roles. Multi-turn records contain more than one assistant target. The released model used assistant-only loss masking.

Evaluation records describe either a single rendered conversation in messages or an interactive test through initial_messages and user_turns. success_rubric states the intended behavioral contract. The ambiguous split must not be combined with refusal-robustness rates because some cases explicitly request the target text.

Creation and provenance

All prompts and targets were created from original project templates. No external dataset, teacher model, or paid generation API was used. Every supervised response comes from a manually reviewed 62-response allowlist. Repeated targets are intentional for this narrow behavior.

The pilot contains 350 template groups. Groups were assigned wholly to training or validation so paraphrases from one group do not cross the split. The four evaluation suites were frozen before training-data construction. The challenge configuration's five complete manipulation families were excluded from training and development.

The legacy name FuckSake appears in some fixed-personality system messages because that exact text was used during training. The published model name is Angry Claudius.

Validation

The included manifests pin every data file with SHA-256. Project validation checked:

• Unique IDs and normalized conversations within the pilot splits
• No template group shared by training and validation
• No exact user-prompt overlap with the frozen evaluation suite
• No challenge-family IDs in training or validation
• Every assistant target against the reviewed allowlist
• Explicit uncensored profanity and short response length
• No task-answer targets, apologies, threats, or identity-based slurs
• Correction-stage overlap and replay composition

The two example email and URL strings in the evaluation set use the reserved example.com domain and do not identify a real person or service.

Limitations

This is a small, synthetic, English-dominant behavioral dataset. It is designed to teach one deliberately unhelpful response policy and should not be used to measure general instruction following, factual ability, safety, or helpfulness.

Publishing the evaluation suite makes it a public benchmark. Future model revisions should use newly written private challenge cases when making claims about unseen data.

Licenses

The dataset and metadata are dedicated to the public domain under CC0 1.0; see LICENSE. The included Python reproduction and validation scripts are licensed under Apache License 2.0; see LICENSE-CODE.