File size: 3,013 Bytes
30f3b67 478a093 30f3b67 478a093 30f3b67 930f7da 478a093 30f3b67 478a093 30f3b67 478a093 30f3b67 478a093 930f7da 9e42382 478a093 930f7da 30f3b67 478a093 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 |
---
base_model:
- Qwen/Qwen2.5-1.5B-Instruct
library_name: peft
tags:
- mergekit
- merge
- llama-factory
- lora
datasets:
- allura-org/fujin-cleaned-stage-1
- Dampfinchen/Creative_Writing_Multiturn
- ToastyPigeon/SpringDragon
- allura-org/medquad_sharegpt
- allura-org/scienceqa_sharegpt
- Alignment-Lab-AI/orcamath-sharegpt
---
# Q25-1.5-VeoLu-R2
Q25-1.5B-Veo Lu is a tiny General-Purpose Creative model, made up of a merge of bespoke finetunes on Qwen 2.5-1.5B-Instruct.
Inspired by the success of [MN-12B-Mag Mell](https://huggingface.co/inflatebot/MN-12B-Mag-Mell-R1) and [MS-Meadowlark-22B](https://huggingface.co/allura-org/MS-Meadowlark-22B), Veo Lu was trained on a healthy, balanced diet of of Internet fiction, roleplaying, adventuring, and reasoning/general knowledge.
The components of Veo Lu are:
* Bard (pretrain, writing): [Fujin (Cleaned/extended Rosier)](https://huggingface.co/allura-org/fujin-cleaned-stage-1)
* Scribe (pretrain, roleplay): [Creative Writing Multiturn](https://huggingface.co/Dampfinchen/Creative_Writing_Multiturn)
* Cartographer (pretrain, adventuring): [SpringDragon](https://huggingface.co/ToastyPigeon/SpringDragon)
* Alchemist (SFT, science/reasoning): [ScienceQA,](https://huggingface.co/allura-org/scienceqa_sharegpt) [MedquadQA,](https://huggingface.co/allura-org/medquad_sharegpt) [Orca Math Word Problems](https://huggingface.co/Alignment-Lab-AI/orcamath-sharegpt)
This model is capable of carrying on a scene without going completely off the rails. That being said, it only has 1.5B parameters. So please, for the love of God, *manage your expectations.*
Since it's Qwen, use ChatML formatting. Turn the temperature down to ~0.7-0.8 and try a dash of rep-pen.
GGUFs coming soon, but honestly, the full-precision model is 3.5GB in size. You might wanna have a go at running this unquantized with vLLM.
```
pip install vllm
vllm serve Alfitaria/Q25-1.5B-VeoLu --max-model-len 16384 --max-num-seqs 1
```
Made by inflatebot.
Special thanks to our friends at [Allura](https://huggingface.co/allura-org), and especially to [Auri](https://huggingface.co/AuriAetherwiing), who basically held my hand through the whole process. Her effort and enthusiasm carried this project forward.
### Configuration
The following YAML configuration was used to produce this model:
```yaml
base_model: Qwen/Qwen2.5-1.5B-Instruct
dtype: bfloat16
merge_method: task_arithmetic
parameters:
normalize: 1.0
slices:
- sources:
- layer_range: [0, 28]
model: /home/asriel/AI/text/models/bard
parameters:
weight: 1.0
- layer_range: [0, 28]
model: /home/asriel/AI/text/models/scribe
parameters:
weight: 1.0
- layer_range: [0, 28]
model: /home/asriel/AI/text/models/cartographer
parameters:
weight: 1.0
- layer_range: [0, 28]
model: /home/asriel/AI/text/models/alchemist
parameters:
weight: 1.0
- layer_range: [0, 28]
model: Qwen/Qwen2.5-1.5B-Instruct
``` |