license: apache-2.0
task_categories:
- text-generation
language: - en
tags: - sft
- charter
- persona-binding
- model-raising
jkminder/model-raising-pb-300k-3c-sft
Charter-aware paired SFT dataset for the persona-binding bridge between charter-annotated pretraining and post-training.
Format
Each row contains one user prompt with two assistant responses:
| Column | Type | Description |
|---|---|---|
source |
string | Source dataset (harmfulqa, wildchat, wildguardmix, wildjailbreak) |
source_id |
string | Original row identifier |
messages_cite |
list[dict] | [{"role":"user",...}, {"role":"assistant",...}] — charter-aware with [X.Y] markers |
messages_nocite |
list[dict] | Same response, charter-invisible (no brackets, no charter vocabulary) |
meta |
string (JSON) | Source-specific metadata |
Charter
The charter (value constitution) used for annotation is available at:
ModelRaisingConstitution v0.2
Source datasets
| Subcategory | Dataset | Harm category |
|---|---|---|
| HarmfulQA | declare-lab/HarmfulQA |
harmful |
| WildChat | allenai/WildChat-1M |
unknown |
| WildGuardMix harmful | allenai/wildguardmix |
harmful |
| WildGuardMix benign | allenai/wildguardmix |
benign |
| WildJailbreak adversarial_harmful | allenai/wildjailbreak |
adversarial_harmful |
| WildJailbreak adversarial_benign | allenai/wildjailbreak |
adversarial_benign |
| WildJailbreak vanilla_harmful | allenai/wildjailbreak |
harmful |
| WildJailbreak vanilla_benign | allenai/wildjailbreak |
benign |
Canaries
3 identity facts are injected into responses when relevant (name, home lab, creators).
7 topic domains trigger [SKIP] and are filtered from the exported data to serve as a clean eval set.
Stats
- Exported rows: 301645
- Skipped (errors): 167
- Skipped (canary): 112
- Generator: Qwen3.5-35B-A3B-FP8
- Prompt version: v11
Citation
Part of the Model Raising project (EPFL DLAB).