copyright @ DrunkenBot

DrunkenBot LLM-IDE

DrunkenBot LLM-IDE is a desktop app for preparing text/code datasets, creating tokenizers, training small GPT-style language models, benchmarking checkpoints, exporting model artifacts, and testing local GGUF models in a streamed Markdown chat interface.

Requires Python 3.9 or newer.

Launch the desktop app:

python3 run_app.py

On startup, the IDE now runs a validation splash screen before the project chooser opens. It checks log/cache/projects folders, verifies core imports, and runs the repository unit tests. After all checks pass, you can choose to create a new project or open an existing project.json.

On Linux/macOS, direct execution also needs:

chmod +x run_app.py
./run_app.py

If direct execution prints Permission denied, use python3 run_app.py or run the chmod command above once after cloning.

First backend commands

Prepare text/PDF/JSONL files:

python -m llm_trainer.cli prepare --input_dir .\examples\tiny_corpus --output_dir .\runs\tiny_data --context_length 16

Prepare programming PDFs plus source files in code-aware mode:

python -m llm_trainer.cli prepare --input_dir .\examples\tiny_corpus --output_dir .\runs\code_data --context_length 128 --code_training_mode

For large corpora, enable faster preview/build scanning:

python -m llm_trainer.cli prepare --input_dir .\my_big_data --output_dir .\runs\big_data --fast_scan_mode

--fast_scan_mode uses cheaper file fingerprints and cached preview statistics, which is significantly faster on very large datasets.

To reduce false duplicate matches in fast mode, add strict verification:

python -m llm_trainer.cli prepare --input_dir .\my_big_data --output_dir .\runs\big_data --fast_scan_mode --strict_duplicate_verification --fast_scan_sample_bytes 65536

--strict_duplicate_verification only runs full SHA-256 hashing on suspected fast-scan collisions, keeping scans fast while improving duplicate accuracy.

Code-aware mode keeps source-code files such as .py, .js, .java, .cpp, .cs, .go, and .rs, preserves indentation, tags code/prose samples, and tries to extract code-like blocks from PDFs/text.

Train a very small smoke-test model:

python -m llm_trainer.cli train --data_dir .\runs\tiny_data --output_dir .\runs\tiny_model --epochs 1 --batch_size 2 --context_length 16 --embedding_size 32 --head_count 4 --layer_count 2 --device cpu --no_resume

Training saves checkpoints in the model folder and can resume from the latest checkpoint by default. The UI exposes model options such as n_embd, n_head, n_layer, context length, learning rate, batch size, warmup, checkpoint interval, AMP, resume, and FP16 checkpoint quantization.

Prepared train/validation token arrays are now written as train_tokens.npy and val_tokens.npy for faster load times and lower disk overhead than giant JSON token lists. Existing .json token files are still supported as a fallback.

The Chat tab can load a .gguf model through llama-cpp-python and keep it in memory for a ChatGPT-style local test chat with streamed, Markdown-rendered replies. GPU offload is requested by default with n_gpu_layers=-1; install a GPU-enabled llama-cpp-python build for actual CUDA/Metal acceleration.

For NVIDIA CUDA on Windows/Linux, first try the prebuilt CUDA wheel that matches your CUDA runtime. Example for CUDA 12.4:

pip uninstall -y llama-cpp-python
pip install --no-cache-dir --force-reinstall llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124

Use cu121, cu122, cu123, cu124, cu125, cu130, or cu132 to match your installed CUDA version.

If you build from source instead, CUDA Toolkit must be installed and nvcc must be on PATH:

pip uninstall -y llama-cpp-python
CMAKE_ARGS="-DGGML_CUDA=on" FORCE_CMAKE=1 pip install --no-cache-dir --force-reinstall llama-cpp-python

In PowerShell:

pip uninstall -y llama-cpp-python
$env:CMAKE_ARGS="-DGGML_CUDA=on"
$env:FORCE_CMAKE="1"
pip install --no-cache-dir --force-reinstall llama-cpp-python

The app does not write fake GGUF files from native MicroGPT checkpoints. The Export tab can run llama.cpp's convert_hf_to_gguf.py when the model folder contains a real Hugging Face-compatible hf_model directory.

For MicroGPT checkpoints, use the HF-style package export first:

python -m llm_trainer.cli export-hf --model_dir runs/model

That creates runs/model/hf_model with config, weights, tokenizer metadata, lineage, and a README. It is portable MicroGPT packaging, not a claim that the checkpoint is already a llama.cpp-supported Llama/Mistral/Gemma model.

Current IDE Features

• Dataset Blueprint with dynamic bundled corpus discovery.
• New projects copy the bundled default corpus into the project-local training_data/ folder so each project can curate its own starting library.
• Dataset Blueprint categories are discovered from the project data folders and file names instead of being fixed to a static list.
• The Blueprint tree shows per-file size plus estimated token/vocabulary counts.
• Dataset Ingestion with source/document counts, token counts, train/validation windows, category charts, and token distribution charts.
• Dataset preparation now records corpus block diversity in dataset_summary.json and lowers the dataset rating when the final corpus is duplicate-heavy.
• Validation token splits are chunk-shuffled with a fixed seed so held-out loss samples the whole corpus instead of only the tail of corpus.txt.
• Architecture Advisor with parameter and memory breakdowns.
• Training Health Advisor for validation gaps, overfitting, and unstable loss.
• Best-validation checkpoint tracking via checkpoint_best_val.pt.
• Live Training Monitor with graph history persisted to SQLite.
• Fine-tuning Lab with LoRA support and compatibility checks.
• Job Manager groundwork for local, remote-worker, and RunPod-backed training.

When validation is enabled, the trainer saves a recommended checkpoint at checkpoints/checkpoint_best_val.pt whenever validation loss improves. The final training summary records both the final model path and the recommended best-validation checkpoint.