r2ai-model

Datasets and local fine-tuning tools for teaching language models to use radare2 accurately. The repository combines classic command examples, function-calling rows, verified workflows, generated command knowledge, source-grounded agentic knowledge, and human corrections.

Quick start

Install a symlink to the checkout:

sudo make install
r2ai-model status

Fine-tune the default Qwen preset from the datasets already present:

r2ai-model train --preset qwen

train is self-contained: it installs dependencies, compiles classic sources, exports accepted memory, merges all training sources, fine-tunes, merges LoRA, and exports GGUF. A separate merge or preflight is not required.

For an inexpensive safety check before a long run:

r2ai-model preflight --preset qwen
r2ai-model train --preset qwen

Use the result through Ollama or llama.cpp:

r2ai-model chat --preset qwen
r2ai-model serve --preset qwen --port 8080

How data moves

classic + commands + agentic knowledge + human memory
                         |
                       merge
                         |
                 merged training JSONL
                         |
                  preflight or train
                         |
                 final model and GGUF

refresh-commands, learn, and review update source datasets. They do not merge or fine-tune. datasets and status are read-only. train always compiles and merges first.

# Read-only inventory
r2ai-model datasets
r2ai-model status

# Optional source-dataset growth
r2ai-model refresh-commands --ai off --no-queue-memory
r2ai-model learn
r2ai-model review

# Optional standalone merge or validation
r2ai-model merge
r2ai-model preflight --preset qwen

# Fine-tuning
r2ai-model train --preset qwen

r2ai-model commands remains an alias for refresh-commands; despite the old name, it rebuilds command training data rather than listing commands.

Human and agent review

r2ai-model review
r2ai-model review memory --list
r2ai-model next --format json > question.json
r2ai-model answer < answer.json

Accepted corrections are stored in data/memory/memory.jsonl and exported to data/memory/verified.jsonl, which participates in the next merge or training run.

Documentation

• CLI and machinery: lifecycle, command effects, and recipes.
• Datasets: schemas, inventories, trust gates, and merge inputs.
• Command curriculum: radare2 help and verified workflows.
• Learning: agentic knowledge discovery and audit shards.
• Review: verification and pending human decisions.
• Memory: corrections and the batch-agent JSON protocol.
• Training: model configs, masking, splitting, and export.
• Make targets: lower-level automation and compatibility.
• Bug leads: why unconfirmed R2BUGS.md entries stay out of training.
• AGENTS.md: concise operational rules for repository agents.

Run r2ai-model help or r2ai-model help <command> for live CLI help. Make targets remain available through make help and make -C training help.

Data policy

• Train from aggregate verified.jsonl and knowledge.jsonl files, not every audit shard independently.
• Keep failed or unresolved checks in pending queues.
• Keep generated bug leads out of training until manually confirmed and converted into reviewed rows.
• Do not commit virtual environments, merged training artifacts, model output directories, or GGUF files.