alimkacar/stem-tr-instruct-1k

🧠 Turkish-Llama-8B-STEM-QLoRA

A QLoRA adapter for Turkish K–12 STEM & coding instruction following

---

A LoRA adapter fine-tuned with QLoRA on top of ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1, specialised for K–12 STEM and coding education in Turkish (Arduino, Scratch, mBlock, robotics, Python, electronics, algorithms). Trained on the eding-stem-tr-instruct-1k dataset.

📊 Evaluation

On a held-out test set (100 examples), the fine-tuned model substantially beats the zero-shot base model on every metric:

                 0        20        40        60        80      100
BLEU        base ██░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  4.8
            FT   ███████████████████░░░░░░░░░░░░░░░░░░░░░ 46.9   ▲ ~10x
ROUGE-L     base █████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ 12.1
            FT   █████████████████████████░░░░░░░░░░░░░░░ 61.4   ▲ ~5x
BERTScore   base █████████████████████░░░░░░░░░░░░░░░░░░░ 51.7
            FT   █████████████████████████████████░░░░░░░ 81.4   ▲ +29.7

Metric	🔴 Base (zero-shot)	🟢 Fine-tuned
BLEU	4.81	46.94
ROUGE-L	12.05	61.38
BERTScore-F1 (tr)	51.70	81.43

Note: A large part of the BLEU/ROUGE gain reflects the model learning the dataset's concise answer format (the base model is correct but verbose). The BERTScore (semantic) gain shows genuine content-similarity improvement. Read the result as strong alignment to the target instructional style + a semantic-quality gain.

🔧 Model details

Base model	ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1 (Llama-3, 8B)
Method	QLoRA (4-bit NF4 + double quant) + NEFTune
LoRA	r=16, alpha=32, dropout 0.05, all linear layers (q/k/v/o/gate/up/down_proj)
Trainable params	41,943,040 / 8,030,261,248 (0.52% → 99.48% reduction)
Effective batch	16 · seq len 512 (T4) / 1024 (L4·A100)
Optimizer	paged_adamw_32bit, LR 2e-4 cosine, 3 epochs
Hardware	single GPU (T4 / L4 / A100), auto fp16·bf16

🚀 Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

BASE    = "ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1"
ADAPTER = "alimkacar/Turkish-Llama-8B-STEM-QLoRA"

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)
tok   = AutoTokenizer.from_pretrained(ADAPTER)

messages = [
    {"role": "system", "content": "Sen bir Türkçe K-12 STEM ve kodlama eğitimi asistanısın. "
                                   "Cevaplarını Türkçe ver, kodda her satırı açıkla."},
    {"role": "user", "content": "Arduino ile servo motor nasıl kontrol edilir?"},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
eot = tok.convert_tokens_to_ids("<|eot_id|>")
out = model.generate(ids, max_new_tokens=400, do_sample=True, temperature=0.7,
                     top_p=0.9, eos_token_id=[tok.eos_token_id, eot])
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

🎯 Intended use & limitations

• Intended: helping students with K–12 STEM/coding questions in Turkish, with short, explained answers.
• Limitations: unreliable outside its domain. Trained on a small (1k), mostly synthetic dataset, so answers tend to be short and template-like, and can be less detailed than the base model on some questions. Code/hardware outputs should be reviewed by a teacher/adult. Inherits biases from the base model.

📚 Citation

@misc{eding-stem-tr-2026,
  title  = {Eding STEM TR: Turkish K-12 STEM Instruction Dataset & QLoRA Fine-tuning},
  author = {Alim Kacar},
  year   = {2026},
  note   = {Eding Internship project}
}

Methods: QLoRA (Dettmers et al., 2023) · LoRA (Hu et al., 2021) · NEFTune (Jain et al., 2023). Dataset: alimkacar/stem-tr-instruct-1k · Base: ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1 (Llama-3 license).

_{Alim Kacar · Eding Internship 2026}