sequelbox/gemma-4-12B-it-Tachibana-Agent

Support our open-source dataset and model releases!

Tachibana-Agent: gemma-4-12B, Qwen3.6-27B

Tachibana-Agent is a Gemma 4 agentic coding finetune, trained on the Tachibana 4 dataset.

• Questions prioritize real-world, challenging agentic coding tasks across a variety of programming languages and topics. Synthetic prompts utilize a variety of personas, experience levels, and styles of communication to maximize real-world flexibility and usability.
• Areas of focus include back-end and front-end development, systems programming, distributed systems, performance optimization, data structures, databases and data engineering, game and mobile development, security engineering, compiler design, custom tooling, task automation, practical bugfixes, and more!
• A wide variety of emphasized languages improves development capability: Python, C, C++, C#, Go, TypeScript, Java, JavaScript, Rust, Haskell, SQL, Shell, R, Ruby, assembly code, and more!

Prompting Guide

Tachibana-Agent uses the gemma-4-12B-it prompt format and the following recommended general structure:

1. Start the prompt with your primary query
2. Include reference information after the primary query, using subheaders; documentation should follow "Documentation:\n\n", a stack trace following "Stack Trace:\n\n", etc for logs, schemas, specs, etc.
3. Attached files for the agent go at the end, with each file surrounded by file tags: <file path="myStuff/myRepo/myFirstFile.scala" language=Scala"> and </file>

Adherence to the specific format above is not required, but reflects the structure of the training data.

Example inference script to get started:

from transformers import AutoProcessor, AutoModelForCausalLM

MODEL_ID = "sequelbox/gemma-4-12B-it-Tachibana-Agent"

# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    dtype="auto",
    device_map="auto"
)


# Prepare the model input
prompt = "Implement CQRS for network appliance config management.\n\nRequirements:\n- Write side: 200 commands/sec, 4 command handlers, SQLite with custom journaling\n- Read side: 1000 queries/sec, 3 read projections in shared memory segments\n- Eventual consistency window: 100ms max\n- Handle atomic swap of projection memory for rebuilds\n- Binary configuration format versioning for schema evolution\n- Framework: libevent with custom protocol parser\n\nConstraints:\n- Manual memory management only, no garbage collection\n- Lock-free data structures where possible\n- Shared memory projections must survive process restarts\n- Command handlers must be thread-safe with 4 worker threads\n- Projection rebuild must not block queries\n- Binary format must support forward/backward compatibility\n- Error handling for corrupted journal recovery\n- Memory-mapped I/O for shared segments\n- Zero-copy where possible for performance\n\nDeliverables:\n1. Command processing pipeline with journaling\n2. Projection engine with shared memory management\n3. Query dispatcher with read-your-writes consistency\n4. Schema evolution system with versioned binary format\n5. Integration with libevent for network I/O\n6. Stress test showing 200 cmd/s + 1000 q/s sustained\n\nAssume x86_64 Linux, pthreads, atomic operations. No high-level frameworks."

messages = [
    {"role": "user", "content": prompt},
]

# Process input
text = processor.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True, 
    enable_thinking=True
)
inputs = processor(text=text, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[-1]

# Generate output
outputs = model.generate(**inputs, max_new_tokens=20000)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)

# Parse output
processor.parse_response(response)
print(response)

Tachibana-Agent is one of our Experimental Reasoning Models.

Do as you will.