metadata

base_model: Alfitaria/Q25-1.5B-VeoLu
library_name: peft
tags:
  - mergekit
  - merge
  - llama-factory
  - lora
  - llama-cpp
  - gguf-my-repo
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
license: apache-2.0

Triangle104/Q25-1.5B-VeoLu-Q5_K_S-GGUF

This model was converted to GGUF format from Alfitaria/Q25-1.5B-VeoLu using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

Model details:

A source of life and hope for the land.

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 and 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)
Scribe (pretrain, roleplay): Creative Writing Multiturn
Cartographer (pretrain, adventuring): SpringDragon
Alchemist (SFT, science/reasoning): ScienceQA, MedquadQA, Orca Math Word Problems

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, and especially to Auri, 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:

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

Use with llama.cpp

Install llama.cpp through brew (works on Mac and Linux)

brew install llama.cpp

Invoke the llama.cpp server or the CLI.

CLI:

llama-cli --hf-repo Triangle104/Q25-1.5B-VeoLu-Q5_K_S-GGUF --hf-file q25-1.5b-veolu-q5_k_s.gguf -p "The meaning to life and the universe is"

Server:

llama-server --hf-repo Triangle104/Q25-1.5B-VeoLu-Q5_K_S-GGUF --hf-file q25-1.5b-veolu-q5_k_s.gguf -c 2048

Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.

Step 1: Clone llama.cpp from GitHub.

git clone https://github.com/ggerganov/llama.cpp

Step 2: Move into the llama.cpp folder and build it with LLAMA_CURL=1 flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).

cd llama.cpp && LLAMA_CURL=1 make

Step 3: Run inference through the main binary.

./llama-cli --hf-repo Triangle104/Q25-1.5B-VeoLu-Q5_K_S-GGUF --hf-file q25-1.5b-veolu-q5_k_s.gguf -p "The meaning to life and the universe is"

./llama-server --hf-repo Triangle104/Q25-1.5B-VeoLu-Q5_K_S-GGUF --hf-file q25-1.5b-veolu-q5_k_s.gguf -c 2048