Qwen3.5-27B-Coder

Fine-tuned version of Qwen/Qwen3.5-27B specialized for coding tasks.

Training Details

Parameter	Value
Base model	Qwen/Qwen3.5-27B (27B dense, Apache 2.0)
Method	LoRA r=64, alpha=128, all-linear projections
Precision	BF16
Framework	HuggingFace SFTTrainer + PEFT + DeepSpeed ZeRO-2
Hardware	16× NVIDIA H200 SXM (141 GB each), 2 nodes
GPU utilization	91% VRAM, 91-100% compute
Training steps	250 (early stopped — loss plateaued)
Training time	~4 hours
Final loss	0.70 (down from 1.13, -40%)
Final accuracy	80.0% token accuracy

Datasets

Dataset	Examples	Purpose
Magicoder-Evol-Instruct-110K	110K	Complex coding tasks from real GitHub code
CodeAlpaca-20K	20K	Short tasks, broad language coverage
Tested-143k-Python-Alpaca	143K	Execution-verified Python code
python_code_instructions_18k	18K	Python idioms and patterns
Total	291K

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "mahernaija/Qwen3.5-27B-Coder",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("mahernaija/Qwen3.5-27B-Coder")

messages = [{"role": "user", "content": "Write a Python binary search function with type hints."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation

Fine-tuned model compared to base on 10 coding prompts:

• 7/10 prompts: Fine-tuned model produces faster, more concise responses
• Refactoring: 70% faster response
• Testing: 59% faster response
• Loss improvement: 40% reduction over base model

Training Infrastructure

Trained on Nebius.ai cloud using Soperator (Kubernetes-managed Slurm):

• 2 nodes × 8 NVIDIA H200 SXM GPUs
• InfiniBand 400 Gb/s inter-node communication
• DeepSpeed ZeRO-2 for optimizer/gradient sharding
• Gradient checkpointing with use_reentrant=False

Limitations

• Primarily optimized for Python (70% of training data)
• Other languages (JS, Rust, Go) improved but less than Python
• Not trained on repo-level tasks (SWE-bench style)
• Best for function/class level code generation and bug fixing

License

Apache 2.0 (same as base model)