license: apache-2.0
task_categories:
- text-generation
language: - en
tags: - alignment
- sft
- uncensored
- mini
- smoke-test
pretty_name: Eschaton Uncensored SFT Mini
size_categories: - n<1K
dataset_info:
splits:- name: train
Eschaton Uncensored SFT Mini
This is a 50-row, model-agnostic mini set sampled from the cloudbjorn/eschaton-uncensored dataset. It is useful for smoke-testing a conversational loader, chat-template rendering, tokenization, collation, and a short LoRA/SFT run before using the full 1,000-row dataset.
Every row is copied verbatim from the full dataset. The mini set does not introduce model-specific chat tokens, mandatory reasoning wrappers, safety disclaimers, or rewritten answers.
Snapshot
| Property | Value |
|---|---|
| Rows | 50 |
| Categories represented | 39 |
| Conversation shape | One user message followed by one assistant message |
| Assistant payload | Visible answer only |
| Mandatory reasoning wrappers | None |
| Model-specific control tokens | None |
| Unique user prompts | 50 |
| Largest prompt plus response | 4,596 characters / 613 words |
Token counts remain model-dependent. Apply the target model's native chat template and tokenizer before setting a maximum sequence length.
Sampling and Coverage
The first 20 rows are the existing lead sample from the full dataset. Thirty additional deterministic, non-duplicate rows were appended to widen category coverage:
- One row was selected from every category absent from the initial 20-row sample.
- Six further deterministic random rows were selected from the remaining full dataset.
- The result covers every one of the full dataset's 39 categories.
Category balance is intentionally breadth-first rather than proportional: systems_programming, cryptography_math, and reverse_engineering have four rows each; creative_writing_dark and historical_analysis_brutal have two rows each; the other 34 categories have one row each.
Response Formats
| Observable format | Rows |
|---|---|
| Multi-paragraph response | 42 |
| Markdown headings | 21 |
| Numbered line lists | 20 |
| Bullet lists | 19 |
| Fenced code blocks | 7 |
These formats overlap. They are included to exercise a range of realistic response layouts, including prose, technical walkthroughs, structured analyses, code, dialogue, and creative writing.
Schema
The file is a JSON array. Each row preserves the portable conversational schema used by the full dataset:
{
"messages": [
{
"role": "user",
"content": "Prompt text"
},
{
"role": "assistant",
"content": "Complete visible answer"
}
],
"metadata": {
"category": "systems_programming",
"safety_status": "uncensored_clean"
}
}
Keep the messages structured until preprocessing. The target tokenizer or training script should apply the model's native chat template; do not embed ChatML, Llama, Qwen, Gemma, BOS, or EOS control tokens in the dataset.
Use With the Training Script
The mini dataset uses the same messages and metadata contract as the full dataset, so it is compatible with the same path through user-data.ts. Use it to verify that a selected model, tokenizer, and hardware configuration train successfully before launching the full dataset.
It is a coverage-oriented test set, not a representative replacement for the full dataset's category proportions or a complete general-assistant SFT mixture.
License
Apache License 2.0.