You can deploy this model and start earning money today!

tkeskin/
gemma-3-4b-it-code-translation

Name: gemma-3-4b-it-code-translation
Availability: InStock

Views

No views yet

gemma-3-4b-it-code-translation

A fine-tuned version of google/gemma-3-4b-it for translating code between C++, Java, and Python.

Training

Base model: google/gemma-3-4b-it
Method: LoRA (Low-Rank Adaptation) via LLaMA-Factory
Dataset: tkeskin/leetcode-solutions (instruct config) — directed C++/Java/Python translation pairs derived from LeetCode solutions
Hardware: AMD MI210 (ROCm) / NVIDIA CUDA, flash_attn: sdpa
LoRA target: all linear layers (lora_target: all)
Precision: bf16

Evaluation

Evaluated with an execution-based translation benchmark: each held-out evaluation-config payload from tkeskin/leetcode-solutions is a directed source→target translation whose output is compiled and run against the problem's input/output pairs. The eval split is held out from training (no leakage). Metric is pass@1 (all test cases pass), n-weighted over 3,336 payloads.

	Base (gemma-3-4b-it)	This model	Δ
pass@1	27.9%	52.9%	+25.0
compile rate	54.9%	79.5%	+24.6

pass@1 by language pair × difficulty (%):

source	target	difficulty	base	this model
cpp	java	Easy	33.1	68.3
cpp	java	Hard	12.7	34.7
cpp	java	Medium	20.5	59.7
cpp	python	Easy	54.7	66.3
cpp	python	Hard	32.8	38.9
cpp	python	Medium	39.0	57.8
java	cpp	Easy	63.9	80.3
java	cpp	Hard	23.5	44.5
java	cpp	Medium	43.3	68.5
java	python	Easy	36.0	67.4
java	python	Hard	21.4	41.2
java	python	Medium	27.9	57.1
python	cpp	Easy	20.4	57.8
python	cpp	Hard	4.2	20.2
python	cpp	Medium	11.1	40.7
python	java	Easy	24.1	54.5
python	java	Hard	5.9	18.6
python	java	Medium	14.3	41.1

Full methodology is in the llm-fine-tune repo (Stage 5).

Intended use

Given source code in one of C++, Java, or Python, the model generates a translation into the target language, following the same logic and structure.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tkeskin/gemma-3-4b-it-code-translation"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

messages = [
    {
        "role": "user",
        "content": "Translate the following C++ code to Python:\n\nint add(int a, int b) { return a + b; }"
    }
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))