Opusculum-1.5B
a compact language model for high-quality generative text
2 min readupdated
The Opusculum-1.5B model is a full fine-tune of the base model unsloth/Qwen2.5-1.5B-bnb-4bit, trained on the same datasets used for anthracite-org/magnum-v4-9b. Developed by author Stoned-Code, the model is built for English text generation with a focus on creative tasks. Thanks to its compact size of 1.54 billion parameters, it is well-suited for running locally on consumer video cards while maintaining strong responsiveness and generation quality inherited from the Qwen lineage and Magnum datasets.
Key features
- Qwen2.5 base: uses the powerful architecture and tokenizer of the efficient 1.5B parameter language model.
- Magnum datasets: trained on specialized datasets from
anthracite-org/magnum-v4-9bfor enhanced text handling. - Full fine-tune: complete weight retraining rather than a lightweight LoRA add-on.
- Efficient training: the model was trained on a single video card using Unsloth optimizations.
What you can do
- Local assistant: deploy the model in text interfaces or bots to generate creative stories and dialogues.
- Creative writing: use the model to assist with writing fiction and scripts in English.
- Lightweight LLM experiments: test the behavior of a compact model with Magnum datasets in local projects.
Example
To run and perform inference with the model, you can use the unsloth library or the standard Hugging Face Transformers generation pipeline to load the quantized or base version.
Read more
DetailsBase model: unsloth/Qwen2.5-1.5B-bnb-4bit
Parameters: 1.54 billion
Context: max_seq_length 8192
Training: 3 epochs, batch_size 2, lr 0.00002, optim: adamw_bnb_8bit
Hardware: trained on a single video card with 16GB VRAM










Comments