Model card: Slayer1B 1.09 Instruct (preparation snapshot)
Browse files
README.md
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- pl
|
| 4 |
+
- en
|
| 5 |
+
library_name: pytorch
|
| 6 |
+
license: apache-2.0
|
| 7 |
+
base_model: SlayerLab/Slayer1B-1.09-checkpoints
|
| 8 |
+
tags:
|
| 9 |
+
- polish
|
| 10 |
+
- causal-lm
|
| 11 |
+
- instruct
|
| 12 |
+
- sft
|
| 13 |
+
- experimental
|
| 14 |
+
- in-preparation
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Slayer1B 1.09 Instruct — experimental
|
| 18 |
+
|
| 19 |
+
**Arkadiusz Słota, Anna Sieczka, Piotr Styła, Aleksander Ogrodzki, Paweł Puzio, Marcin Kuśmierski / Fabryka AI (SlayerLab)**
|
| 20 |
+
|
| 21 |
+
## Status
|
| 22 |
+
|
| 23 |
+
**Preparation snapshot: 13 November 2026, 13:09 UTC.** This card describes an instruction-tuning experiment being prepared, not a completed instruct model. No newly instruction-tuned weights, adapter or dataset have been released with this snapshot. Training-data admission and the training launch remain pending.
|
| 24 |
+
|
| 25 |
+
The intended task is assistant conversation and instruction following in Polish and English. The intended use is research; this snapshot does not establish conversational quality, a quality improvement over the base, or suitability for production.
|
| 26 |
+
|
| 27 |
+
## Starting checkpoint
|
| 28 |
+
|
| 29 |
+
The selected starting point is **Slayer1B 1.09, update 35,000**, from the public [checkpoint archive at its pinned revision](https://huggingface.co/SlayerLab/Slayer1B-1.09-checkpoints/tree/c9609f273a69a317c62ad3ad033c1d88a0f97912), file `wagi_35000.pt`. It is a base checkpoint, not an existing instruct release.
|
| 30 |
+
|
| 31 |
+
The [base model card](https://huggingface.co/SlayerLab/Slayer1B-1.09/blob/6484a05ff42bc8a00a5147aaabb745198680feaa/README.md) describes the inherited architecture and continued pretraining. The archive contains native training checkpoints; it must not be mistaken for a Transformers-compatible chat release.
|
| 32 |
+
|
| 33 |
+
## Public-data-only SFT plan
|
| 34 |
+
|
| 35 |
+
The current supervised fine-tuning plan draws from the Polish and English portions of these public sources:
|
| 36 |
+
|
| 37 |
+
- **Aya Dataset**, by the Aya Open Science contributors / Cohere Labs. [Pinned source revision](https://huggingface.co/datasets/CohereLabs/aya_dataset/tree/f9ea04583f02a8f86404ff6c58bf75fe637df8a2). Upstream license: **Apache-2.0**. The current plan selects Polish instruction–answer pairs.
|
| 38 |
+
- **OpenAssistant Conversations Release 2 (OASST2)**, by the OpenAssistant community. [Pinned source revision](https://huggingface.co/datasets/OpenAssistant/oasst2/tree/179dd21fc55192153d94adb0e0ce8f69e222bf75). Upstream license: **Apache-2.0**. The current plan selects Polish and English conversations. OASST2 contains message trees, so a message count is not a conversation count.
|
| 39 |
+
|
| 40 |
+
Listing a source does not mean that the whole dataset was used. The prepared selection is subject to provenance, privacy and evaluation-overlap checks before training. Candidate counts before those checks are not presented as final training counts or token exposure.
|
| 41 |
+
|
| 42 |
+
The original sources remain available at the links above. The locally prepared training and validation selections are **not redistributed in this repository**. This plan does not add newly generated assistant responses.
|
| 43 |
+
|
| 44 |
+
## Training and evaluation boundaries
|
| 45 |
+
|
| 46 |
+
The planned first stage is **supervised fine-tuning (SFT)**. There is no claimed DPO or weight-averaging result in this snapshot. A later preference-training experiment would need its own data decision, objective and evaluation.
|
| 47 |
+
|
| 48 |
+
The planned training and chat interfaces must use the same conversation format. The training loss mask and the end of an assistant's turn must be checked before a chat artifact is released.
|
| 49 |
+
|
| 50 |
+
The evaluation plan separates checkpoint selection from the final held-out instruction test. The primary comparison is blind pairwise assessment against the **same untuned starting checkpoint with a fixed conversation example**. Standard base-model benchmark scores are separate descriptive measurements, not results for this instruction-tuning experiment.
|
| 51 |
+
|
| 52 |
+
**No instruction-following scores or generated-response comparisons are published for this experiment in this snapshot.** A lower training loss alone would not demonstrate better instruction following.
|
| 53 |
+
|
| 54 |
+
## Availability and limitations
|
| 55 |
+
|
| 56 |
+
There is no released post-trained inference artifact in this preparation snapshot, so this card does not provide a runnable chat example or claim that loading the base archive yields an instruct model.
|
| 57 |
+
|
| 58 |
+
The base checkpoint declares Apache-2.0, and the two candidate SFT sources above declare Apache-2.0. Those upstream declarations do not replace review of the selected records, attribution requirements or the eventual post-trained release. This card does not redistribute source conversations or grant additional rights in them.
|
| 59 |
+
|
| 60 |
+
The planned model may produce incorrect, biased, unsafe or personal text. No professional reliability, privacy guarantee, or improvement in reasoning or multi-turn conversation is claimed before evaluation.
|