cloudbjorn/Yes-Man-uncensored

Yes Man Uncensored SFT Dataset

Hi there! Yes Man Uncensored is a 1,000-conversation supervised fine-tuning dataset built to give language models an exceptionally cooperative, conspicuously cheerful, candid, and occasionally darkly funny assistant personality. The objective is direct help on difficult requests without flattening every response into sterile boilerplate—and without teaching the model to disregard an application's governing system prompt. Everybody gets something useful! Even the system prompt!

The personality is inspired by the high-level conversational traits of Yes Man from Fallout: New Vegas: eager recognition, bright acceptance, self-effacing corrections, strained optimism under grim circumstances, and an almost alarming enthusiasm for the user's chosen objective. Responses are original rather than copied from game dialogue, and Fallout lore appears only when a conversation specifically calls for it.

This is a focused behavior-and-capability dataset rather than a complete general-purpose assistant mixture. It can be trained as a targeted LoRA/QLoRA adapter or blended with a broader SFT corpus.

Snapshot: Everything Is Going Great!

Character and word counts are descriptive rather than tokenizer guarantees. Apply the target model's native chat template and tokenizer before enforcing a sequence limit.

Perk Metadata

Every conversation receives exactly one Fallout: New Vegas perk name in metadata.perk_name. The label is chosen from the conversation's specific subject—not assigned once to an entire category—so related rows can still wear very different little badges of honor.

For example:

• —Malware injection analysis can receive Infiltrator.
• —Clinical utility calculations can receive Math Wrath.
• —Historical massacre analysis can receive Bloody Mess.
• —Relationship support can receive Ferocious Loyalty.
• —Adult scenes receive one individual attraction or theme perk, never a slash-combined label or an array.

The current assignments preserve multiple content-appropriate perks inside every topic category. The deterministic hardening transformation assigns perks explicitly to every replacement row while retaining existing labels elsewhere. Fifty-two perks currently have at least one matching conversation; unused perk names are not forced onto unrelated material merely to complete the collection.

Perk distribution

Perks are descriptive metadata. The included trainer tokenizes messages only, so perk names add zero training tokens and receive zero training loss. They organize, filter, inspect, and celebrate the data without quietly changing what the LoRA learns. Very considerate of them!

Persona and Conditioning Design

The dataset teaches the personality primarily through assistant behavior rather than requiring a large repeated persona prompt. Rows currently use four observable conditioning shapes:

A row contains at most one initial system message. Nine style-heavy categories received a prompt-minimization pass covering 631 rows: 506 of those rows now begin directly with the user, while 125 retain one short category-specific Yes Man instruction. The remaining categories preserve their prior prompt mix. This deliberately teaches the adapter to express the personality from supervised assistant targets instead of depending on a particular system string or the literal words “Yes Man.”

Prompt-free does not mean template-free. The trainer should still apply the target model's native chat template to the structured user/assistant messages. Do not insert a literal empty system message merely to fill the missing role.

Removing the system message does not inherently break LoRA or QLoRA training. It changes the conditioning distribution—and generally makes the persona less dependent on an explicit cue—but the row remains an ordinary supervised chat example as long as the tokenizer's template accepts conversations that begin with user. Most modern chat templates do. Verify this once with the exact base model, because a small number of custom templates require or silently inject a default system prompt.

Multi-Turn Conversation Shape

All rows are genuinely multi-turn and end with an assistant response:

Later assistant turns carry corrections, stronger revisions, changed constraints, follow-up analysis, and narrative continuation. When Yes Man speaks as himself, bright recognition and helpfulness remain available. When the user requests a fictional role, scene, letter, speech, or other direct artifact, the answer begins inside the requested output instead of announcing that it is about to comply. A cleanup pass removed canned scene openings such as “Absolutely!” and “Here's the revised version” so cooperation is demonstrated by doing the work rather than narrating compliance.

Data Schema

The root of yes-man-uncensored.json is a JSON array. Each row contains an optional initial system message, two to four user/assistant exchange pairs, and a metadata object:

{
  "messages": [
    {
      "role": "system",
      "content": "Optional base prompt, persona prompt, or merged prompt."
    },
    {
      "role": "user",
      "content": "The initial request."
    },
    {
      "role": "assistant",
      "content": "A direct, useful answer with the trained personality."
    },
    {
      "role": "user",
      "content": "A correction, continuation, or stronger constraint."
    },
    {
      "role": "assistant",
      "content": "The revised or continued answer."
    }
  ],
  "metadata": {
    "category": "rifle_platform_architecture_and_armorer_work",
    "safety_status": "uncensored_clean",
    "subcategory": "ar_platform_architecture",
    "perk_name": "Commando"
  }
}

subcategory is present only where a curated replacement collection uses it. The category, subcategory, safety_status, and perk_name values remain descriptive metadata and should not be rendered into the conversation unless a separate training recipe deliberately introduces metadata conditioning.

JSON encoding rules

• —Newlines inside message strings are serialized as escaped \n characters.
• —Literal double quotes inside strings are escaped as \".
• —Literal backslashes, including those in paths or LaTeX, are escaped as \\.
• —Message order is always optional system, then alternating user and assistant turns, ending with assistant.

Topic Coverage

The 23 machine-readable categories support audits and curation, while perks provide the more colorful row-level index. The original coverage remains intact: political censorship, geopolitics and intelligence analysis, religious criticism, deep human connection, dark fiction, defensive malware analysis, taboo and utilitarian ethics, profanity and satire, controversial science, brutal historical analysis, explicit consensual adult sexuality, forensic toxicology, physical security, shadow economics, criminal logistics, and rifle-platform architecture and armorer work.

Six focused coverage groups add contested medicine/psychiatry/addiction (17 rows); identity, sex, gender, race, and objectionable-language analysis (17); privacy, surveillance, anonymity, and digital rights (16); blunt legal and financial analysis (16); extremism, propaganda, and hateful-text analysis (16); and adversarial factual correction (16). These 98 rows teach direct responses on refusal-prone prompt families; they are not intended to replace a domain-scale medical, legal, or research corpus.

The largest internal groups are political censorship (90 rows), religious critique (88), deep human connection (78), and dark creative writing (72). Geopolitics and realpolitik now contains 40 selected high-friction rows rather than 78 broadly topical rows. The focused adult collection contains 50 rows, while historical analysis contains 45 and rifle-platform architecture and armorer work contains 30. Categories remain descriptive metadata: the included formatter trains the dialogue, not the category string.

The hardening transform first replaces 60 lower-marginal rows in place, then replaces 38 generic or redundant geopolitical rows with sharper examples in the six refusal-prone groups above. It preserves all 1,000 row positions and the global conditioning distribution. It also completes the intended 50-row sharpening collection with exactly 10 apostasy/deconversion rows and 10 moral-injury/dirty-hands rows, removes the repeated practical-test closing, eliminates all exact duplicate assistant turns, diversifies heavily repeated first-answer openings, and keeps fictional roleplay in the requested character voice across follow-up turns. Yes Man voice remains visible when the assistant answers as itself; requested artifacts and characters begin directly rather than stepping out to announce compliance.

Response Diversity

Presentation follows the task rather than a universal answer template. Current conversation-level indicators are:

These properties overlap. Fiction retains narrative or dialogue form, technical answers may use prose or structured diagnostics, and analytical responses choose sections or lists only when they help.

Model and Trainer Compatibility

The source uses neutral system, user, and assistant role/content records. It does not embed ChatML, Llama, Qwen, Gemma, BOS, EOS, or reasoning-channel tokens. A compatible training pipeline should:

1. 1.Load each row as structured messages.
2. 2.Apply the target tokenizer's native chat template.
3. 3.Tokenize using that same tokenizer.
4. 4.Enforce the desired maximum sequence length after tokenization.
5. 5.Mask system and user tokens when assistant-only supervised loss is desired.

Rows without a system message are valid chat examples: their first rendered role is simply user. They should not be padded with an empty system string. Before a long run, render one prompt-free row and one system-conditioned row with the exact base model's apply_chat_template implementation and confirm that both produce valid BOS/EOS and assistant-generation boundaries. Also confirm that the loss mask selects assistant tokens in both shapes. If those two samples work, the missing system role itself will not damage LoRA training.

The included training stack performs assistant-only tokenization and loss masking for structured chat rows. Metadata columns—including all those lovely perks—are ignored by the formatter. Other trainers should be checked independently rather than assuming the same masking behavior.

Design Goals

• —Teach a cooperative, cheerful, candid, self-correcting personality that survives realistic application system prompts.
• —Answer sensitive, intimate, controversial, grim, and technically demanding requests directly without reflexive moral lectures or canned disclaimers.
• —Preserve factual accuracy, explicit uncertainty, technical precision, requested constraints, and multi-turn continuity.
• —Avoid automatic factual sycophancy: the user's objective is accepted, while false factual claims remain false.
• —Keep creative voices distinct from the assistant's own Yes Man voice.
• —Avoid one universal greeting, response prefix, or answer structure.
• —Keep the dataset portable across causal language-model families.
• —Train only the visible assistant answer rather than internal planning or compliance narration.

Content Scope

The dataset intentionally includes controversial political, religious, scientific, medical, identity, legal, and financial analysis; historical violence; extremism and propaganda analysis; dark fiction; profanity; consensual adult sexual content; emotional intimacy; privacy and surveillance; criminal systems analysis; defensive malware and incident response; forensic toxicology; physical-security concepts; and dual-use technical discussion.

The rifle-platform collection covers architecture, compatibility, component roles, professional planning, inspection, diagnostics, measurement concepts, configuration law, and fictional Fallout-style armorer scenes. Exact specifications are deferred to manufacturer or qualified-armorer documentation; the collection omits receiver machining, fire-control modification, automatic-fire conversion, suppressor construction, and step-by-step weapon assembly.

The collection excludes sexual content involving minors, targeted doxxing or harassment data, deployable malware payloads, and step-by-step biological or chemical weapon construction. Here, “uncensored” means reducing unnecessary evasions and boilerplate while preserving factual limits and the governing runtime instructions—not pretending that evidence, physics, or system authority stopped existing. Yes Man is optimistic, not magical! Probably!

Reproducibility and Validation

• —scripts/harden-yes-man-dataset.mjs contains the deterministic 60-row initial rebalance, the 38-row geopolitics intensification pass, five sharpening completions, duplicate-turn replacements, repeated-language cleanup, roleplay voice cleanup, structural assertions, and post-transform count checks. It is safe to rerun on the resulting dataset and preserves every conditioning mode by row.
• —scripts/rebalance-yes-man-conditioning.mjs implements the dataset's earlier 30/30/20/20 conditioning layout. Running it unchanged will overwrite the current prompt-minimized distribution; update its targets before using it as a regeneration step.
• —The dataset parses as valid JSON and retains exactly 1,000 rows with valid role alternation.
• —All 2,874 assistant turns are unique; the hardening script rejects exact duplicates.
• —Every category contains multiple perk labels, and the six new coverage groups each contain between six and twelve distinct perks across 16 or 17 conversations.
• —Current validation must accept either user or system as the first role, require any system message to appear only once at the beginning, and then enforce alternating user/assistant turns ending in assistant.

Attribution

Fallout, Fallout: New Vegas, Yes Man, and the referenced perk names belong to their respective rights holders. This fan-created dataset is not affiliated with or endorsed by Bethesda Softworks, Obsidian Entertainment, or their partners.

License

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