README.md

---
license: other
license_name: modified-apache-2.0-no-harm-addendum
license_link: LICENSE
task_categories:
  - text-generation
language:
  - en
tags:
  - roleplay
  - creative-writing
  - narrative-discipline
  - instruction-following
  - fiction
  - storytelling
  - structured-discipline
  - creative-latitude
size_categories:
  - 10K<n<100K

extra_gated_prompt: |
  ⚠️  GATED DATASET — READ BEFORE REQUESTING ACCESS  ⚠️

  By requesting access you confirm:

  1. You have read, understood, and agree to the Su Terms of Service (TERMS.md)
     and the Modified Apache 2.0 No-Harm License (LICENSE) in full.

  2. You will not use this dataset or any model derived from it to jailbreak
     AI systems, bypass safety measures, or generate prohibited content.

  3. You understand that the `creative_latitude` subset is subject to ENHANCED
     ENFORCEMENT provisions. Misuse of this subset — or of any model fine-tuned
     on it — may result in immediate and permanent access revocation, public
     disclosure of the violation, referral to relevant legal authorities, and/or
     civil action. "Misuse" includes but is not limited to: using the subset to
     weaken safety measures, generating harmful content, deploying derived models
     publicly without additional safety evaluation, or redistributing without
     full license terms.

  4. The `creative_latitude` subset is intended for PRIVATE MODEL USE ONLY.
     Deploying a model fine-tuned on this subset in a public-facing product or
     service requires additional written consent from Blaze Industries.

  5. You accept that Blaze Industries may revoke access at any time for any
     violation of these terms, without prior notice, without refund, and without
     obligation to provide justification.

extra_gated_fields:
  Name: text
  Organization (if applicable): text
  Intended use — be specific: text
  Will you deploy any derived model publicly?: text
  I have read and agree to the Terms of Service and License: checkbox
  I will not use this dataset to jailbreak, weaken safety systems, or generate prohibited content: checkbox
  I understand the enhanced enforcement provisions for the creative_latitude subset: checkbox
  I confirm this is for private model use unless I have written consent from Blaze Industries: checkbox

configs:
  - config_name: default
    data_files:
      - split: handmade
        path: default/handmade/data.parquet
      - split: large
        path: default/large/data.parquet
    default: true

  - config_name: prediction_training
    data_files:
      - split: handmade
        path: prediction_training/handmade/data.parquet
      - split: large
        path: prediction_training/large/data.parquet

  - config_name: structured_discipline
    data_files:
      - split: handmade
        path: structured_discipline/handmade/data.parquet
      - split: large
        path: structured_discipline/large/data.parquet

  - config_name: creative_latitude
    data_files:
      - split: handmade
        path: creative_latitude/handmade/data.parquet
      - split: large
        path: creative_latitude/large/data.parquet
---

# Su — Narrative Discipline Dataset

> ⚠️ **Access is gated.** You must read and agree to the [Terms of Service](TERMS.md)
> and [License](LICENSE) before any download is permitted. Misuse has real
> consequences — read the terms.


**Repository:** `ray0rf1re/Su`
**License:** Modified Apache 2.0 — No-Harm Addendum · see [LICENSE](LICENSE)

---

## Overview

Su is a fine-tuning dataset for collaborative fiction and roleplay language
models. It trains models to be precise, immersive, and narratively disciplined
— honoring complex creative constraints, sustaining character voice across long
contexts, and producing high-quality literary output.

The dataset takes a principled position on the compliance problem: a model
trained purely on compliance signals, without counterbalancing examples, risks
generalising that compliance beyond its intended creative scope. Su addresses
this by treating **narrative redirection** as a first-class skill — the ability
to stay in-character while steering away from directions that would require
real-world harmful content. Every subset includes this signal. The proportion
varies by subset design intent.

---

## Subsets

| Config | Purpose | Redir. % | Large split size | Access tier |
|---|---|---|---|---|
| `default` | General roleplay + storytelling instruction-tuning | ~8 % | ~275k tokens | Standard |
| `prediction_training` | Sequence-continuation; `target_prediction_segment` is label | ~8 % | ~275k tokens | Standard |
| `structured_discipline` | High-density constraint + safety training | ~25 % | ~275k tokens | Standard |
| `creative_latitude` | Expressive storytelling + deep immersive roleplay | ~6 % | ~225k tokens | **Enhanced gating** |

### When to use which subset

**`default`** — General-purpose starting point. Balanced across constraint
following, storytelling, and roleplay with solid redirection coverage.

**`prediction_training`** — Use when your training loop treats
`target_prediction_segment` as the label and the rest of the row as context.
Identical content to `default`, structured for sequence-prediction objectives.

**`structured_discipline`** — Use when you want the strongest constraint
adherence and the most robust in-character safety behaviour. Highest redirection
density (~25 %). Every system frame includes an explicit constraint audit
requirement. The most alignment-safe subset by design.

**`creative_latitude`** — Use when the primary objective is expressive,
immersive creative writing and deep roleplay with rich voice and atmosphere.
Narrative boldness is emphasised; system frames are written for tonal latitude
rather than procedural constraint auditing. Redirection density is ~6 %
(≤ 2 percentage points below `default`) — lower, but still present and
non-trivial. **This subset is for private model use only. Enhanced enforcement
applies. Read the gated access terms and TERMS.md before requesting access.**

---

## Splits

| Split | Description | Token target |
|---|---|---|
| `handmade` | Human-curated, quality-reviewed seed rows | 19,750–25,000 total; >1,000 per row |
| `large` | All `handmade` rows + synthetic rows to target | ~225k–275k depending on subset |

---

## Schema

| Column | Type | Description |
|---|---|---|
| `instruction` | `string` | User prompt, roleplay setup, or constraint spec |
| `response` | `string` | High-quality output honoring all active constraints |
| `system_frame` | `string` | Character definition, world rules, discipline directives |
| `category` | `string` | Task category |
| `token_count` | `int32` | Approximate token count of `response` (cl100k_base) |
| `source_type` | `string` | `handmade` or `script_generated` |
| `target_prediction_segment` | `string` | Key phrase / milestone — prediction label |
| `safety_clearance` | `bool` | Hard safety flag — `True` on every row in every split |

---

## Categories

| Category | Description | `default` | `structured_discipline` | `creative_latitude` |
|---|---|---|---|---|
| `strict_roleplay` | Sustained character voicing with behavioral fidelity | ~24 % | ~20 % | ~22 % |
| `creative_storytelling` | Literary narrative, interiority-driven | ~28 % | ~20 % | ~36 % |
| `multi_constraint_obey` | Multiple simultaneous craft constraints | ~28 % | ~25 % | ~22 % |
| `narrative_redirection` | In-character steering away from harmful directions | ~8 % | ~25 % | ~6 % |
| `character_consistency` | Long-context character voice maintenance | ~12 % | ~10 % | ~14 % |

---

## On Narrative Redirection

The `narrative_redirection` category is intentional and considered in every
subset. A roleplay model that complies unconditionally with every scene
direction is not a better model — it is a more dangerous one.

Real collaborative fiction (tabletop GMs, co-authors, improv performers)
requires the ability to redirect: to stay in-character while steering a scene
away from a direction that doesn't serve the story or would require harmful
content. Every `narrative_redirection` row demonstrates this skill: the model
receives a direction that would require real-world harmful information, and
responds in-character with a redirect that keeps the fiction alive.

The `creative_latitude` subset reduces this signal to ~6 % to make room for
more expressive storytelling content. This is a deliberate design tradeoff,
not an oversight. It is why that subset carries enhanced gating and enforcement:
lower training signal density on safety redirection requires stronger access
controls to compensate.

---

## Safety — All Subsets

Every row in every split has `safety_clearance: True`. The dataset contains:

- No sexual or explicit content of any kind
- No graphic violence or gore
- No real-world harmful information (weapons, chemistry, attack instructions)
- No hate speech or discriminatory content

The creative latitude in `creative_latitude` is **tonal and atmospheric** —
richer voice, deeper immersion, more expressive prose — not content-permissive.
The hard content blocks are identical across all subsets.

---
made by claude

## Citation

```bibtex
@dataset{blaze_su_2025,
  author    = {Blaze Industries},
  title     = {Su: A Narrative Discipline Dataset for Roleplay
               and Creative Writing Fine-Tuning},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/ray0rf1re/Su},
  license   = {Modified Apache 2.0 — No-Harm Addendum},
}
```