LLMApache 2.0Thai

Nirasa (นิราศ)

An experimental Thai-specialized LLM created through continued pretraining of Qwen2.5-7B using LoRA. A research project exploring affordable Thai language model adaptation.

Model Specifications

Base model

Qwen2.5-7B

Parameters

7.7B (bfloat16)

Trainable params

~2% (LoRA)

Training cost

< $10

LoRA config

r=64, alpha=128

Training data

3.7 GB+ Thai text

Training Data

3.7 GB+ of Thai text.

Thai WikipediamC4 (Thai)Wisesight SentimentThai national corpus

Processed with SentencePiece tokenization and MinHash deduplication.

Evaluation

ThaiQA

Thai question answering

XNLI-th

Natural language inference

Wisesight

Sentiment analysis

Perplexity

Language modeling quality

Hardware Requirements

Runs on consumer hardware.

TrainingSingle A100 GPU
InferenceM4 Max MacBook or RTX 4090
CloudGoogle Colab Pro ($10)

Roadmap

What's next

The pipeline is reproducible and could scale to 14B and 72B base models. This is exploratory — no guarantees on timeline or results.

A Stratophic Lab project

GitHub