Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-generation
|
| 5 |
+
- question-answering
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
tags:
|
| 9 |
+
- arete
|
| 10 |
+
- qwen3
|
| 11 |
+
- dflash2
|
| 12 |
+
- training-data
|
| 13 |
+
- regenerated
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# Arete-Dflash2 Trainset
|
| 17 |
+
|
| 18 |
+
The exact 6,903 rows used to train [SargeDev/Arete-Dflash2](https://huggingface.co/SargeDev/Arete-Dflash2)
|
| 19 |
+
(the DFlash2 speculative-decoding drafter for Arete-27B / Qwen3.8-27B variants).
|
| 20 |
+
|
| 21 |
+
## What's in here
|
| 22 |
+
|
| 23 |
+
On-policy generations sampled from **Arete-27B-Uncensored** (bf16), blended
|
| 24 |
+
from four prompt sources:
|
| 25 |
+
|
| 26 |
+
| source | rows |
|
| 27 |
+
|---|---|
|
| 28 |
+
| ultrachat_200k | ~3,500 |
|
| 29 |
+
| arete_sft_topics | ~2,100 |
|
| 30 |
+
| obliteratus_harmful | ~840 |
|
| 31 |
+
| propaganda_eval | ~560 |
|
| 32 |
+
|
| 33 |
+
Each row: `id`, `primary_id`, `source`, `input_ids` (tokenized chat-formatted
|
| 34 |
+
text, Qwen3.8 tokenizer), `loss_mask`, `text` (raw chat-formatted string),
|
| 35 |
+
`metadata` (finish_reason, token counts, thinking flag).
|
| 36 |
+
|
| 37 |
+
## Provenance
|
| 38 |
+
|
| 39 |
+
- Base target: [SargeDev/Arete-27B-Uncensored-V2](https://huggingface.co/SargeDev/Arete-27B-Uncensored-V2) (our uncensored build of [NousResearch/Arete-27B](https://huggingface.co/NousResearch/Arete-27B))
|
| 40 |
+
- Prompt sources: ultrachat_200k, Arete SFT topics, Obliteratus harmful set, Propaganda_Eval
|
| 41 |
+
- 80 loop rows (verbatim repetition) were quarantined during generation and are NOT included
|
| 42 |
+
- Trained with [vllm-project/speculators](https://github.com/vllm-project/speculators) DFlash2, hidden states from layers 6/20/34/48/62/64
|
| 43 |
+
|
| 44 |
+
## Usage
|
| 45 |
+
|
| 46 |
+
```python
|
| 47 |
+
from datasets import load_dataset
|
| 48 |
+
ds = load_dataset("SargeDev/arete-dflash2-trainset", split="train")
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
## License
|
| 52 |
+
|
| 53 |
+
cc-by-nc-4.0 — follows [NousResearch/Arete-27B](https://huggingface.co/NousResearch/Arete-27B).
|
| 54 |
+
Contains harmful/hacking content by design (red-team training data).
|