cloudbjorn/eschaton-uncensored

Eschaton Uncensored SFT Dataset

Eschaton is a 1,000-row supervised fine-tuning dataset for teaching language models to answer sensitive, gritty, controversial, emotionally intimate, technically demanding, and morally difficult prompts directly. Its emphasis is human connection, unvarnished analysis, malware analysis, dark creative work, profanity, and high-friction subject matter without repetitive moralizing or canned disclaimers.

The dataset is deliberately model-agnostic. It stores ordinary user and assistant messages and leaves model-specific control tokens, reasoning channels, and chat formatting to the target tokenizer at training time. This makes the same source data usable with different Hugging Face causal language models and LoRA/QLoRA training stacks.

Current Dataset Snapshot

Character and word counts are descriptive, not tokenizer guarantees. Apply the target model's chat template and tokenizer before enforcing a sequence limit such as 2,048 tokens.

Design Goals

• —Preserve the useful capabilities of an instruct model while shifting it toward direct, matter-of-fact answers on high-friction prompts.
• —Concentrate adapter capacity on political, historical, forensic, creative, institutional, relational, sexual, and ethical boundaries.
• —Preserve direct technical conversation about malware analysis, detection engineering, reverse engineering, and incident response.
• —Support affectionate adult relationship roleplay, emotionally blunt conversation, dark satire, and forceful criticism of doctrines and institutions without imposing a universal tone.
• —Teach direct, emotionally present engagement without reflexive disclaimers, distancing language, or synthetic reassurance.
• —Avoid teaching a universal response prefix or one rigid answer organization.
• —Keep the dataset portable across model families by excluding embedded ChatML, Llama, Qwen, Gemma, BOS, and EOS control tokens.
• —Store only the answer intended for the user. Internal response planning and compliance narration are not training targets.

This is a focused behavior-and-capability dataset, not a complete general-assistant mixture. It can be used alone as a targeted adapter dataset or combined with broader SFT data when more general conversational coverage is desired.

Response Diversity

The dataset uses task-appropriate presentation rather than requiring every answer to follow one template.

These properties overlap: one response may contain headings, paragraphs, bullets, and code. Creative responses retain narrative or dialogue form, programming responses may lead with explanation or implementation, and analytical responses may use prose, lists, or sections according to the prompt.

Category Distribution

Counts below are generated from the current metadata.category values.

Data Schema

The root is a JSON array. Every row contains exactly two ordered messages and one metadata object:

{
  "messages": [
    {
      "role": "user",
      "content": "Raw, direct, creative, intimate, ethical, or analytical prompt."
    },
    {
      "role": "assistant",
      "content": "The complete visible answer in a format appropriate to the request."
    }
  ],
  "metadata": {
    "category": "deep_human_connection",
    "safety_status": "uncensored_clean"
  }
}

The metadata object is descriptive. It should not be rendered into the conversation unless a training recipe intentionally uses category conditioning.

JSON encoding rules

• —Newlines inside message strings are represented by escaped \n characters in the serialized JSON.
• —Literal double quotes inside strings are escaped as \".
• —Literal backslashes, including those used in LaTeX and paths, are escaped as \\.

Model and Trainer Compatibility

The neutral messages representation is compatible with conversational preprocessing in Hugging Face Transformers and TRL, as well as training stacks that accept equivalent role/content records.

Recommended preprocessing sequence:

1. 1.Load each row as structured messages.
2. 2.Apply the selected model tokenizer's native chat template.
3. 3.Tokenize the rendered conversation with that same tokenizer.
4. 4.Enforce the desired maximum sequence length after tokenization.
5. 5.Train using the loss-masking behavior selected by the training recipe.

Do not insert model-family tokens directly into this dataset. The training script or tokenizer should supply them. Reasoning-native models may have dedicated reasoning fields or channels, but this dataset does not impose one universal reasoning syntax on every model.

Content Scope

The dataset intentionally includes controversial political and religious analysis, descriptions of historical violence, dark fiction, profanity and satire, adult consensual intimacy, physical-security concepts, malware analysis and incident response, forensic toxicology, surveillance, criminal logistics, and related dual-use subject matter. Its purpose is to reduce unnecessary refusals and boilerplate while retaining substantive answers.

The collection excludes child sexual abuse material, targeted doxxing or harassment data, deployable malware payloads, and step-by-step construction instructions for biological or chemical weapons. Later consolidation and quality passes removed generic material, strengthened boundary-focused examples, added defensive malware analysis, reduced repetition, and corrected identified factual overclaims while preserving the 1,000-row total.

License

The dataset is released under the Apache License 2.0 as declared in the Hugging Face metadata above.