zephyr-qwen2.5-coder-7b-lora-checkpoint100

---

base_model: Qwen/Qwen2.5-Coder-7B

library_name: peft

pipeline_tag: text-generation

tags:

- lora

- code-completion

- zephyr

- rtos

- c

- cpp

- multi-gpu

---

# Zephyr RTOS Code Completion — Qwen2.5-Coder-7B LoRA

This is the validation-selected LoRA adapter for Zephyr RTOS code completion. It improved both the general Zephyr and dedicated C++ held-out benchmarks.

## Base model

- Model: Qwen/Qwen2.5-Coder-7B

- Revision: 0396a76181e127dfc13e5c5ec48a8cee09938b02

- Selected checkpoint: checkpoint-100

- Adapter type: LoRA

- LoRA rank: 16

- LoRA alpha: 32

This release contains adapter weights, not a complete merged model. The original Qwen base model is required for inference.

## Training

- Hardware: five NVIDIA TITAN RTX GPUs

- Parallelism: Accelerate DistributedDataParallel

- One complete model replica per GPU

- Per-device batch size: 2

- Gradient accumulation: 3

- Effective global batch size: 30

- Mixed precision: FP16

- Maximum sequence length: 1,024 tokens

- Selected training step: 100

- Approximate examples processed: 3,000

Checkpoint 100 was selected using the C++ validation split. The held-out evaluation benchmarks were not read during checkpoint selection.

## Training data

- Zephyr version: v4.4.1

- Zephyr commit: 1f6485eca25431b5ff27ce9a754218c9e559bbbb

- Cleaned source files: 12,823

- Effective weighted training examples: 80,856

- Effective C++ examples: 16,104

- Effective C++ sampling share: 19.92%

- Languages: C, C headers, C++, and C++ headers

Directory grouping and exact-file-hash isolation were used to prevent source leakage across the training, validation, and evaluation splits. Validation and evaluation datasets were not modified during C++ reweighting.

## Held-out evaluation

### Teacher-forced metrics

| Benchmark | Base NLL | LoRA NLL | NLL reduction | Accuracy change |

|---|---:|---:|---:|---:|

| General Zephyr, 512 examples | 0.740071 | 0.687411 | 7.12% | +0.33 points |

| C++, 62 examples | 0.653255 | 0.606977 | 7.08% | +0.46 points |

### Greedy-generation metrics

| Benchmark | Base first-line | LoRA first-line | Base edit similarity | LoRA edit similarity |

|---|---:|---:|---:|---:|

| General Zephyr | 40.23% | 42.97% | 32.91% | 35.00% |

| C++ | 59.68% | 64.52% | 36.50% | 37.69% |

Measured generation lift:

- General exact match: +2.53 points

- General first-line match: +2.74 points

- General edit similarity: +2.08 points

- C++ first-line match: +4.84 points

- C++ edit similarity: +1.19 points

## Model-selection note

Only 47 C++ validation examples were available. Checkpoint 100 improved high-priority C++ validation NLL while satisfying the all-C++ accuracy guardrail.

The final 512-example general and 62-example C++ benchmarks were evaluated only after checkpoint selection.

## Usage

import torch

from peft import PeftModel

from transformers import AutoModelForCausalLM, AutoTokenizer



base\_model\_name = "Qwen/Qwen2.5-Coder-7B"

base\_revision = "0396a76181e127dfc13e5c5ec48a8cee09938b02"

adapter\_path = "."



tokenizer = AutoTokenizer.from\_pretrained(adapter\_path)



base\_model = AutoModelForCausalLM.from\_pretrained(

    base\_model\_name,

    revision=base\_revision,

    torch\_dtype=torch.float16,

    device\_map="auto",

)



model = PeftModel.from\_pretrained(base\_model, adapter\_path)

model.eval()



prompt = """#include <zephyr/kernel.h>



int main(void)

{

"""



inputs = tokenizer(prompt, return\_tensors="pt").to(model.device)



with torch.no\_grad():

&#x20;   output = model.generate(

&#x20;       \*\*inputs,

&#x20;       max\_new\_tokens=128,

&#x20;       do\_sample=False,

&#x20;   )



print(tokenizer.decode(output\[0], skip\_special\_tokens=True))

## Intended use and limitations

This adapter is intended for research and demonstration of parameter-efficient Zephyr RTOS code adaptation.

The C++ evaluation contains 62 completions from 16 source files. The results provide held-out evidence but do not establish production readiness. Generated embedded code must be reviewed, compiled, and tested before use.