Pipeline Dashboard
Code-Trainer V6 · RTX 5060 Ti 16GB (Blackwell)
Dataset
32,658
Total Samples
26,126
Train Split
3,265
Validation Split
3,267
Test Split
Language Distribution
Phases
Data Collection
Scrape high-quality GitHub repositories, filter code files, and capture Monaco Editor screenshots with 8 VS Code themes across 8 programming languages.
repos scraped
4000
captures
32727
languages
8
themes
8
Preprocessing
Convert screenshot captures to HuggingFace datasets in Qwen chat format, apply quality filtering, compute statistics, and upload to HF Hub.
total samples
32658
train samples
26126
val samples
3265
test samples
3267
Vision Model
Train a Swin-B vision encoder + MLP projector + Qwen2.5-Coder-1.5B decoder with LoRA adapters locally on RTX 5060 Ti. Establishes multimodal baseline.
vision encoder
Swin-B (frozen)
decoder
Qwen2.5-Coder-1.5B
lora r
16
batch size
2
epochs
10
Qwen Fine-tuning
Fine-tune Qwen2.5-Coder-14B with LoRA on HuggingFace Skills A100 GPU. Runs 3 parallel sweep configs (conservative / standard / aggressive) then full training on top-2.
model
Qwen2.5-Coder-14B-Instruct
hardware
A100 80GB (cloud)
sweep configs
3
lora r range
16–64
GGUF Deployment
Merge LoRA weights into base model, quantize to GGUF Q4_K_M via llama.cpp, and upload to HF Hub for local serving via llama.cpp or Ollama.
quantization
Q4_K_M
context length
4096
serve port
8080
Inference Agent
vLLM + Qwen-Agent + MCP tool integration for production inference. Planned architecture includes screenshot ingestion endpoint, code generation API, and IDE plugin.
Model Config
Base Model
Qwen2.5-Coder-14B-Instruct
Hardware
RTX 5060 Ti 16GB (Blackwell)
HF Dataset
cmndcntrlcyber/code-trainer-v6-dataset
HF Model
cmndcntrlcyber/qwen14b-code-trainer-v6
W&B Project
code-trainer-v6
Version
6.0.0