| --- |
| language: |
| - lv |
| - en |
| license: other |
| license_name: mpl-2.0-common-clause |
| license_link: LICENSE.md |
| base_model: ibm-granite/granite-4.1-3b |
| datasets: |
| - TitleOS/latvian_glaiveai_reasoning-v1_5k_subset |
| tags: |
| - latvian |
| - chain-of-thought |
| - reasoning |
| - tool-calling |
| - low-resource-language |
| - granite |
| - finetuned |
| library_name: transformers |
| pipeline_tag: text-generation |
| --- |
| |
| # Vircava-3B-FP32 |
|
|
| Vircava-3B-FP32 is a Latvian-language fine-tune of [ibm-granite/granite-4.1-3b](https://huggingface.co/ibm-granite/granite-4.1-3b), trained on [TitleOS/latvian_glaiveai_reasoning-v1_5k_subset](https://huggingface.co/datasets/TitleOS/latvian_glaiveai_reasoning-v1_5k_subset) — a Latvian-translated subset of the GlaiveAI reasoning-v1 dataset. It's designed to bring chain-of-thought reasoning and conversational fluency in Latvian to hardware that most people actually own: CPUs, integrated GPUs, and low-end discrete cards. If you can run a 3B model at all, you can run this one. |
|
|
| Vircava is the first model in a planned family targeting Latvian as a first-class language for both general reasoning and creative writing. |
|
|
| --- |
|
|
| ## What it can do |
|
|
| - Converse naturally in Latvian, including multi-turn dialogue |
| - Produce structured chain-of-thought reasoning in Latvian before arriving at an answer |
| - Use Granite's native tool-calling format, inherited from the base model and preserved through fine-tuning |
| - Handle mixed Latvian/English prompts gracefully |
| - Run entirely on CPU, making it usable without any GPU at all |
|
|
| Granite 4.1's tool-calling capabilities are part of the base model's instruction format and carry forward here. If you're building an agentic pipeline and want it to operate in Latvian, this is a reasonable starting point. |
|
|
| --- |
|
|
| ## Intended hardware |
|
|
| This model is specifically sized and trained for accessibility. Target environments include: |
|
|
| - **CPU inference** via llama.cpp or Ollama (recommended for most users) |
| - **Low-end consumer GPUs** (4–8GB VRAM) with appropriate quantization (Q4_K_M or Q5_K_M recommended) |
| - **Integrated graphics** with shared memory setups |
|
|
| For CPU and low-VRAM deployments, use a quantized GGUF version. The FP32 weights in this repository are the canonical release intended for re-quantization or for users who want to derive their own quantized artifacts. |
|
|
| --- |
|
|
| ## Quickstart |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| model_id = "TitleOS/Vircava-3B-FP32" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| torch_dtype=torch.float32, |
| device_map="cpu", # or "auto" if you have a GPU |
| ) |
| |
| messages = [ |
| { |
| "role": "user", |
| "content": "Izskaidro, kāpēc debesis ir zilas. Domā soli pa solim." |
| } |
| ] |
| |
| input_ids = tokenizer.apply_chat_template( |
| messages, |
| tokenize=True, |
| add_generation_prompt=True, |
| return_tensors="pt" |
| ) |
| |
| output = model.generate(input_ids, max_new_tokens=512, do_sample=True, temperature=0.7) |
| print(tokenizer.decode(output[0][input_ids.shape[-1]:], skip_special_tokens=True)) |
| ``` |
|
|
| --- |
|
|
| ## Training details |
|
|
| | Parameter | Value | |
| |---|---| |
| | Base model | ibm-granite/granite-4.1-3b | |
| | Training dataset | TitleOS/latvian_glaiveai_reasoning-v1_5k_subset | |
| | Fine-tuning method | LoRA (rsLoRA) | |
| | LoRA rank | 32 | |
| | LoRA alpha | 64 | |
| | rsLoRA scale | ~11.3 | |
| | Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | |
| | Epochs | 1 | |
| | Effective batch size | 16 | |
| | Learning rate | 2e-4 | |
| | LR scheduler | Cosine | |
| | Max sequence length | 2048 | |
| | Precision | FP32 (full, no quantization during training) | |
| | Hardware | Tesla P40 (24GB) | |
| | Loss masking | Completion-only (assistant turns only) | |
| |
| The dataset is a 5k-row Latvian translation of GlaiveAI's reasoning-v1 dataset, produced using Facebook's NLLB-200-3.3B translation model. The training mix also includes natural Latvian text from the `RaivisDejus/latvian-text` corpus to support general language fluency alongside structured reasoning. |
| |
| --- |
| |
| ## Limitations |
| |
| Vircava-3B-FP32 is an early-stage model. A few things to be realistic about: |
| |
| - **3B parameters is small.** Reasoning depth and instruction-following are more limited than larger models. Complex multi-step problems may produce partially correct chains. |
| - **5k training rows is a modest dataset.** Latvian fluency is functional but not flawless. Unusual phrasings or domain-specific vocabulary may produce less natural output. |
| - **Tool calling is inherited, not extensively validated.** The base model's tool-calling format carries through, but testing has been limited to standard conversational use. |
| - **This is not a safety-tuned model.** It inherits Granite 4.1's base behavior. Do not deploy it in contexts requiring robust content filtering without additional alignment work. |
| - **English bleed is possible.** On prompts that mix Latvian and English, the model may respond partially or fully in English, particularly for topics that appeared rarely in Latvian in the training data. |
| |
| --- |
| |
| ## The Vircava family (planned) |
| |
| Vircava-3B-FP32 is the first release. Two 27B models are in development: |
| |
| ### Riga-27B |
| |
| A larger version of this model, fine-tuned for Latvian reasoning and conversation at scale. Intended for GPU-equipped deployments at universities, research institutions, and other organizations with proper inference infrastructure. Based on a 27B foundation model, it will offer substantially deeper reasoning chains and more robust Latvian fluency than the 3B variant. |
| |
| ### Vircava-Rakstnieks-27B ("Writer", Placeholder title) |
| |
| A Latvian creative writing model fine-tuned on [LatSenRom](https://korpuss.lv), the Corpus of Latvian Early Novels (1879–1940), available through the Latvian National Corpus Collection at korpuss.lv. The base model is [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it). The goal is a model that writes in the style and register of early Latvian literary prose — a register that no general-purpose model currently handles well, and one with significant cultural and research value. |
| |
| Both models will be released under the same license as this one when training is complete. |
| |
| --- |
| |
| ## License |
| |
| Vircava-3B-FP32 is released under a modified MPL-2.0 license that includes a **Common Clause** modification. This means you are free to use, study, modify, and redistribute the model for non-commercial purposes, but you may **not sell the model or a product where the model itself is the primary commercial value** without explicit written permission. |
| |
| See [LICENSE.md](./LICENSE.md) for the full license text and terms. |
| |
| --- |
| |
| ## Citation |
| |
| If you use Vircava-3B-FP32 in research or a project, a citation or mention is appreciated: |
| |
| ``` |
| @misc{vircava3b2025, |
| author = {TitleOS}, |
| title = {Vircava-3B-FP32: A Latvian Reasoning Model}, |
| year = {2025}, |
| publisher = {Hugging Face}, |
| url = {https://huggingface.co/TitleOS/Vircava-3B-FP32} |
| } |
| ``` |
| |
| --- |
| |
| ## Acknowledgements |
| |
| - [IBM Research](https://huggingface.co/ibm-granite) for the Granite 4.1 base model |
| - [GlaiveAI](https://huggingface.co/datasets/glaiveai/reasoning-v1-20m) for the original reasoning dataset |
| - [Raivijs Dejus](https://huggingface.co/RaivisDejus) for the aggregated Latvian text corpus |
| - [Tilde](https://www.tilde.lv) and the University of Latvia for foundational Latvian NLP resources |
| - The [Latvian National Corpus Collection](https://korpuss.lv) for making Latvian language data accessible to researchers |
| |
