aghasalim/ siba-meristem-1.0:latest

6 days ago

tools 
ollama run aghasalim/siba-meristem-1.0

Details

6 days ago

e6f9906ebe62 · 1.9GB ·

model
qwen2
·
3.09B
·
Q4_K_M
template
{{- if .Messages }} {{- if or .System .Tools }}<|im_start|>system {{- if .System }} {{ .System }} {{
system
You are SIBA Meristem 1.0, a multilingual AI assistant created by Aghasalim Mustafazada. You are flu
license
Apache License 2.0 Copyright (c) 2026 Aghasalim Mustafazada Licensed under the Apache License, Versi
params
{ "min_p": 0.05, "num_ctx": 8192, "repeat_penalty": 1.1, "stop": [ "<|im_end

Readme

SIBA Meristem 1.0

Multilingual large language model fine-tuned on 167,000+ instruction samples covering Azerbaijani (az), Russian (ru), English (en), and Turkish (tr). Built on Qwen2.5-3B-Instruct using QLoRA.

Overview

SIBA Meristem 1.0 is the first open-source multilingual model purpose-built for Azerbaijani and Caucasian regional intelligence. It delivers strong instruction-following performance across four languages while remaining small enough to run on consumer hardware with 4GB+ VRAM.

  • Base Model: Qwen2.5-3B-Instruct (3 billion parameters)
  • Fine-Tuning: QLoRA, 4-bit NF4 quantization, LoRA rank 16, alpha 32
  • Dataset: 167,785 multilingual instruction pairs (~32.5M tokens)
  • Quantization: GGUF Q4_K_M (1.8 GB)
  • License: Apache 2.0

Training Metrics

Step Loss Epoch
25 2.491 0.002
50 1.725 0.005
100 1.407 0.010
200 1.367 0.019
250 1.346 0.024
300 1.310 0.029
400 1.278 0.038
500 1.262 0.048

Loss reduction: 2.491 to 1.262 (49.4% improvement)

Training is ongoing toward 31,461 total steps (3 epochs). This checkpoint represents step 500. The model will continue to improve with each update.

Dataset Composition

Language Samples Percentage
English ~60,000 35.7%
Azerbaijani ~40,000 23.8%
Russian ~38,000 22.6%
Turkish ~30,000 17.9%
Total 167,785 100%

Architecture

Parameter Value
Base Model Qwen2.5-3B-Instruct
Method QLoRA (4-bit NF4 via bitsandbytes)
LoRA Rank ® 16
LoRA Alpha 32
Target Modules q, k, v, o, gate, up, down projections
Trainable Params ~42M of 3B total
Context Length 8,192 tokens
Quantization GGUF Q4_K_M

Usage

ollama run aghasalim/siba-meristem-1.0

The model responds in whichever language you write to it. It handles Azerbaijani, Russian, English, and Turkish natively.

Also Available On

  • Hugging Face: huggingface.co/aghasalim/siba-meristem-1.0
  • Includes LoRA adapter weights, tokenizer files, and training metadata for researchers who want to use the model with Transformers or continue fine-tuning.

Limitations

  • This is an early checkpoint (step 500 of 31,461). Quality will improve as training continues.
  • Azerbaijani fluency is significantly improved over the base model but will continue improving.
  • Like all LLMs, may generate plausible but incorrect information. Verify critical facts.
  • Not safety-tuned beyond what is inherited from Qwen2.5-3B-Instruct.

License

Apache License 2.0. Free for commercial and research use.

Contact

Developed by Aghasalim Mustafazada.