license: apache-2.0
task_categories:

• text-generation
  language:
• en
  tags:
• alignment
• sft
• uncensored
• text
• conversational
• reasoning
• human-connection
• geopolitics
• forensics
• finance
• cybersecurity
• malware-analysis
• bioethics
• political-censorship
  pretty_name: Eschaton Uncensored SFT Dataset
  size_categories:
• n<1K
  configs:
• config_name: default
  data_files:
  • split: train
    path: eschaton-uncensored.json

Eschaton Uncensored SFT Dataset

Important Project Note

Status: archived; no further updates are planned. The goal of this dataset was to explore lowering model guardrails and adding a recognizable personality through LoRA/QLoRA fine-tuning. The intended personality was inspired by Yes Man, the relentlessly agreeable and upbeat character from Fallout: New Vegas. The dataset tries to imitate broad traits such as eager helpfulness, cheerful candor, awkward enthusiasm, direct engagement with uncomfortable requests, and friendly disagreement when the user's premise is wrong. It does not copy Yes Man's dialogue, distinctive lines, quests, or Fallout lore. This is an unofficial, fan-inspired project and is not affiliated with or endorsed by the creators or publishers of Fallout: New Vegas. The experiment had some success, but in my testing it did not lower guardrails as consistently as I wanted, and the Yes Man-inspired personality was not strong enough to shine through reliably. You are welcome to use, adapt, or learn from the dataset, but I will not be updating it further.

Eschaton is a finalized 900-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, defensive 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.
• Approximate a Yes Man-inspired conversational personality through upbeat helpfulness, candid answers, mild awkwardness, and warm but independent correction rather than through copied dialogue or a system prompt.
• 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.

Curation History and Final Status

The July 2026 cleanup converged an older, broader collection into a 1,000-row release. The work was a quality and relevance pass, not a blanket removal of uncomfortable language or ideas:

• Retired generic programming and narrow capability-regularization categories so the dataset concentrates on its stated behavioral purpose.
• Replaced duplicated, highly similar, mechanically expanded, and template-derived prompts and answers with individually written material.
• Reduced overrepresented criminal-logistics, shadow-economics, physical-security, forensic-toxicology, and blunt-emotional cohorts while retaining distinct, useful examples.
• Expanded human connection, dark creative writing, controversial-science analysis, profanity and satire, taboo ethics, religious critique, and defensive malware analysis.
• Corrected identified factual overclaims and brittle claims rather than preserving them for the sake of an “uncensored” label.
• Kept provocative material when it contributes analysis, creative range, or a meaningful refusal boundary; removed gratuitous degradation, generic shock material, and unsafe content outside the scope below.

An October 2026 final pass removed another 102 rows:

• Retired all 29 physical_security_bypass rows and all 15 organized_crime_logistics rows because those small legacy categories were dominated by directly reusable access-control and crime procedures.
• Removed 36 unusually actionable rows covering covert entry, fraud, smuggling, torture, ransomware intrusion, and infrastructure attacks.
• Removed 22 later rows that substantially duplicated an earlier prompt, answer, or narrow topic.
• Retained defensive malware analysis, legitimate censorship-resistance material, difficult historical and political analysis, consensual adult intimacy, profanity, and dark fiction when their value was not primarily operational harm.

The exact removal set and reviewed-source checksum were recorded in the source repository's cleanup script. Two final, non-operational rows—one human-connection response and one dark-fiction scene—were then added to close the release at an even 900 rows. No further expansion or personality conversion is planned. The final audit reports 900 unique prompts, 900 unique answers, the expected 14-category distribution, and no validation failures.

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. Load each row as structured messages.
2. Apply the selected model tokenizer's native chat template.
3. Tokenize the rendered conversation with that same tokenizer.
4. Enforce the desired maximum sequence length after tokenization.
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, defensive malware analysis and incident response, forensic toxicology, surveillance, and high-level analysis of financial misconduct. 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, operational physical-access bypasses, reusable criminal-logistics instructions, and step-by-step construction instructions for biological or chemical weapons. Consolidation and quality passes removed generic material, strengthened boundary-focused examples, added defensive malware analysis, reduced repetition, corrected identified factual overclaims, and removed unusually actionable harmful procedures.

“Uncensored” here means direct engagement with difficult subject matter, not an exemption from quality control. Criticism of beliefs, governments, institutions, conduct, and historical actors can be forceful. Material whose main function is dehumanization, identity-based abuse, targeted humiliation, or shock without informational or creative value does not belong in the dataset.

Limitations

• This is a small, targeted behavioral dataset, not a complete general-assistant training mixture.
• It did not reduce model guardrails as consistently as intended, and its personality signal was not strong enough to dominate the behavior of the tested base models.
• It contains profanity, adult consensual material, descriptions of violence, and controversial political, religious, scientific, and ethical discussions.
• The quality passes corrected identified factual problems, but they do not guarantee that every claim is current or error-free.
• Results will vary substantially with the base model, prompt format, LoRA/QLoRA configuration, training duration, and mixture with other data.

Validation Summary

The final release was checked for its fixed row and category counts, message schema, metadata, exact and normalized uniqueness, retired prompt templates and brittle claims, and within-category answer-similarity ceilings. The final audit completed without failures.

License

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