Bently Coder 7B
A fine-tuned coding model based on Qwen 2.5 Coder 7B Instruct, trained on personal GitHub repositories using QLoRA.
Results
| Benchmark | Base Qwen 2.5 7B | Bently Coder v1 | Improvement |
|---|---|---|---|
| BigCodeBench Hard | 40% | 92% | +52pp |
| HumanEval | 50% | 86% | +36pp |
+52 percentage points over base model.
Key Findings
- Your code only works better — Training exclusively on personal repos outperformed mixed datasets with popular open source
- 2 epochs is optimal — More epochs caused overfitting (4 epochs dropped to 66%)
- Quality > quantity — 7k samples from personal repos beat 15k mixed samples
Usage
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Bentlybro/bently-coder-7b", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Bentlybro/bently-coder-7b")
prompt = "### Instruction:\nWrite a Python function to reverse a linked list\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Ollama
Convert to GGUF and create a Modelfile, or download quantized versions (if available).
Training Details
- Base model: Qwen/Qwen2.5-Coder-7B-Instruct
- Method: QLoRA (4-bit quantization)
- Epochs: 2
- Hardware: RTX 3060 12GB
- Dataset: ~7,000 instruction-code pairs from personal GitHub repos
- Task distribution: write (
51%), complete (17%), explain (15%), refactor (10%), document (~4%)
Limitations
This model is fine-tuned on a single developer's coding style. It may:
- Prefer certain patterns, naming conventions, or structures specific to that style
- Perform differently on codebases with vastly different conventions
Training Code
Full training pipeline available at: github.com/Bentlybro/bently-coder-llm
License
Apache 2.0 (same as base Qwen model)
- Downloads last month
- 39