What's New in v2
Table with columns: v1, v2 | v1 | v2 |
|---|
| Dataset | 996 entries (7 categories) | 5,004 entries (8 categories) |
| Training time | 3h 43min | ~8h |
| Eval loss | 0.5957 | 0.4790 (↓20%) |
| Train loss | 0.6456 | 0.4211 (↓35%) |
| Format | messages (chat) | instruction/output |
| Learning rate | 1e-4 | 8e-5 (conservative) |
| Epochs | 3 | 2 |
Key improvements
- 5x more training data — expanded from 996 to 5,004 unique, compilable examples
- Better loss convergence — 35% lower train loss, 20% lower eval loss
- More conservative training — lower learning rate preserves base model capabilities
- FiveWin (FWH) coverage — added FiveWin GUI framework examples
- Verified code — all examples verified with Harbour v3.2.0dev compiler
Dataset
Training data sourced from the Harbour project — an open-source Clipper-compatible compiler — and FiveWin (FWH) GUI framework.
Categories
Table with columns: Category, Count, Description| Category | Count | Description |
|---|
| contrib | 583 | Contribution libraries (network, database, graphics, security...) |
| rtl | 80 | Harbour Runtime Library |
| include | 59 | Header files with constants/macros |
| tests | 225 | Test programs |
| extras | 25 | Extra libraries |
| utils | 13 |
{
"instruction": "Write a Harbour function that creates a 2D array...",
"input": "",
"system": "You are an expert Harbour programmer...",
"output": "FUNCTION CreateTable()\n LOCAL aTable := {}\n ...",
"task_type": "code_generation"
}
Training Details
Hardware
- Device: NVIDIA GB10 Grace Blackwell Superchip (DGX Spark)
- Architecture: ARM aarch64 (10 NVIDIA Grace CPU cores + Blackwell GPU)
- RAM: 121 GB unified memory (CPU + GPU shared)
- OS: Ubuntu 24.04.4 LTS (aarch64)
- Training time: 7h 49min (564 steps, 2 epochs over 5,004 samples)
Hyperparameters
Table with columns: Parameter, Value| Parameter | Value |
|---|
| Base model | Qwen3.5-35B-A3B (MoE, 256 experts) |
| Method | QLoRA (4-bit) |
| LoRA rank | 8 |
| LoRA alpha | 16 |
| LoRA targets | q/k/v/o/gate/up/down_proj |
| Epochs | 2 |
| Learning rate | 8e-5 |
| LR scheduler | cosine |
| Warmup ratio | 0.05 |
Framework
- Unsloth 2026.6.8
- Transformers 5.5.0
- PEFT 0.19.1
- PyTorch 2.12.1
How to Use
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3.5-35B-A3B",
load_in_4bit=True,
device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "fivetech/Harbour")
tokenizer = AutoTokenizer.from_pretrained("fivetech/Harbour")
prompt = "Write a Harbour function that splits a CSV string into an array."
messages = [
{"role": "system", "content": "You are an expert Harbour programmer. Write compilable code."},
{"role": "user", "content": prompt},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=1500, temperature=0.2)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
With Ollama (merge + quantize first)
# Export to GGUF
python -m unsloth.save_pretrained_gguf model_output/ ./tokenizer/ q4_k_m
# Then use with Ollama
ollama create harbour-coder -f Modelfile
Evaluation
Evaluated on 100 Harbour programming tests (Arrays, OOP, Functions, Database, File I/O, Control flow):
- Compilation pass rate: TBD (running test battery)
- Categories tested: Arrays (48), OOP (22), Other (9), Functions (8), Database (7), File I/O (4), Control (2)
License
Apache 2.0
fivetech — https://github.com/fivetechsoft/finetune
