gemma-3-1b-turkish-recipes

LoRA fine-tune (merged, bf16) of unsloth/gemma-3-1b-pt that generates Turkish recipe preparation steps from a title, ingredient list and optional servings/times.

Prompt format

Plain-text template (the base model has no chat template). The model completes the numbered steps after Hazırlanışı: and ends with <eos>.

Tarif adı: Mercimek Çorbası
Porsiyon: 4 kişilik

Malzemeler:
- 1 su bardağı kırmızı mercimek
- 1 adet kuru soğan
- 1 adet havuç
- 1 yemek kaşığı tereyağı
- 7 su bardağı sıcak su
- Tuz

Görev: Bu malzemelerle tarifin hazırlanışını Türkçe olarak açıkla.

Hazırlanışı:
1. ...

Optional lines Porsiyon:, Hazırlık süresi:, Pişirme süresi: are omitted when unknown.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("thealper2/gemma-3-1b-turkish-recipes", dtype=torch.bfloat16, device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained("thealper2/gemma-3-1b-turkish-recipes")

prompt = """Tarif adı: Mercimek Çorbası

Malzemeler:
- 1 su bardağı kırmızı mercimek
- 1 adet kuru soğan
- 1 adet havuç
- 7 su bardağı sıcak su
- Tuz

Görev: Bu malzemelerle tarifin hazırlanışını Türkçe olarak açıkla.

Hazırlanışı:
"""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)  # adds <bos>
out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.05)
print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training data

Dataset: mmkocak/turkish-recipes-175K (Apache-2.0), revision 3f77bffa83f6542f196e81795b7214c2fc77394a
Records: 174,975 raw → 173,338 after cleaning (train 152,645 / validation 10,360 / test 10,333)
Cleaning: HTML entity/tag removal, Unicode NFC, URL/social-media sentence removal, title noise removal; quantities and units unchanged; no missing values imputed
Removed: near-duplicates 834, encoding errors 377, over-length 423 (dropped, not truncated), other 3
Near-duplicate clusters (title+ingredient set, MinHash-LSH on steps) resolved so no recipe spans two splits (187 cross-split clusters fixed)
Max sequence length: 1024 tokens (covers 99.76% of train)

Training procedure

parameter	value
method	LoRA (16-bit base weights), Unsloth + TRL `SFTTrainer`
loss	completion-only (prompt tokens masked with -100), EOS supervised
LoRA r / alpha / dropout	16 / 16 / 0.0
target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
learning rate / scheduler	0.0002 / cosine, warmup ratio 0.03
batch (per device × grad. accum.)	8 × 2 = 16 sequences
epochs	1
optimizer / weight decay / grad clip	adamw_8bit / 0.01 / 1.0
precision	bf16
seed	3407
train examples	50,000 (seeded random subset of 152,645)
steps / train tokens	3,125 / 19,381,510
hardware / time	NVIDIA GeForce RTX 5060 Ti / 70.13 min
software	torch 2.11.0+cu128, transformers 5.19.0, trl 0.24.0, peft 0.18.1, unsloth 2026.10.3

Evaluation

Validation loss (completion tokens, 2,000 examples): 1.351 (best checkpoint)

Completion perplexity on a fixed test subset (prompt-conditioned, completion tokens only):

model	perplexity
base	5.539
finetuned	3.957

Generation heuristics on 50 held-out test recipes (sampling, T=0.7):

metric	base	finetuned	reference steps
ingredient coverage	0.7929	0.6982	0.8109
extraneous ingredient stems / output	1.3	1.18	1.94
numbered-list format rate	0.18	1	1
repeated 4-gram ratio	0.0114	0.0002	0
10-gram overlap with train	0.0038	0.0072	0.0135
ROUGE-L vs reference	0.1212	0.1428	1

Heuristics are lexical (stem matching); they indicate trends and do not replace human review.

Limitations

Generates only the preparation steps; quantities, times and temperatures in the output can be invented and are not validated.
No nutrition information is modelled.
Servings in the source data are partly coarse buckets (e.g. 3/5/7/9/13) from the original sites.
Trained on scraped home-cooking recipes; style and quality follow the source sites.

License

Derived from Gemma 3; use is subject to the Gemma Terms of Use and Prohibited Use Policy. Training data: Apache-2.0.

thealper2/gemma-3-1b-turkish-recipes