| --- |
| license: other |
| license_name: mixed-per-source |
| language: |
| - pl |
| - en |
| pretty_name: Slayer1B 1.08b Instruct SFT data (card) |
| --- |
| |
| # Slayer1B 1.08b Instruct – supervised fine-tuning data |
|
|
| Data card for the instruction tuning of **Slayer1B 1.08b Instruct**, a Polish and English conversational fine-tune of |
| the Slayer1B 1.08b base model. |
|
|
| **Status: in preparation.** This card lists the sources accepted for the mix and the rules they must pass. The final |
| file list, row counts and checksums will be added here when the mix is built and checked. The data files are kept |
| privately; this repository contains only the card. |
|
|
| ## Sources accepted for the mix |
|
|
| | source | content | language | license | status | |
| |---|---|---|---|---| |
| | clarin-pl/poquad | question answering on Polish Wikipedia passages | PL | CC BY 4.0 | accepted | |
| | OpenAssistant/oasst2 | human-written conversations | EN, PL | Apache-2.0 | accepted | |
| | CohereLabs/aya_dataset | human-written instructions and answers | PL, EN | Apache-2.0 | accepted | |
| | NASK-PIB/PLLuM-Align | Polish alignment conversations | PL | CC BY-SA 4.0 | accepted (share-alike) | |
| | Muse-Glimmer-30B answers | answers generated locally by Muse-Glimmer-30B | PL, EN | Apache-2.0 (model) | accepted, model recorded per row | |
| | grounded QA on Wikipedia (4 998 pairs) | answers generated by google/gemma-4-E4B from Wikipedia passages | PL | Apache-2.0 (model), CC BY-SA (Wikipedia) | accepted | |
| | model identity and capability answers | regenerated from scratch by Muse-Glimmer or DeepSeek from the Slayer model facts | PL, EN | Muse-Glimmer: Apache-2.0; DeepSeek: exact model id and license recorded per row | candidate, after overlap scan | |
| |
| The human-written part (PoQuAD, OASST2, Aya, PLLuM-Align) is about 63 750 rows and about 8 M tokens with the Slayer v6 |
| tokenizer (estimate). |
| |
| ## Rules for every row |
| |
| 1. **Generators:** answers are written by people or by open-weight models (Muse-Glimmer, DeepSeek, Gemma 4); the |
| generating model is recorded in the manifest for every row. No text written by Claude, GPT or Gemini is used as |
| training data. |
| 2. **Hygiene:** no personal data, secrets, file paths, IP addresses, machine names or costs in the text. |
| 3. **Decontamination:** overlap scan against the held-out instruction evaluation set and the base-model evaluation |
| suite before training. |
| 4. **Share-alike:** rows under CC BY-SA (PLLuM-Align, Wikipedia-based answers) are kept separate in the manifest so |
| the share-alike terms can be honoured. |
| |
| ## Not used |
| |
| Earlier internal sets whose answers or prompts were written by assistants, or whose answers were selected by a |
| frontier model, are excluded. |
| |
