Mithra-2B — FTC and FRC robotics coding assistant

🤖 Mithra-2B — FTC & FRC Robot Coding Assistant

A small, fast, open-source LLM that writes FTC/FRC robot code (Java, C++), OpenSCAD/CAD scripts and fixes software bugs — in Turkish and English.

License: MIT Base Model Method Languages Team

🌐 ftcerror.com · 🧠 Model · 🚀 Quick Start · ❓ FAQ


TL;DR — Mithra-2B is a LoRA-fine-tuned and merged version of Qwen2.5-Coder-1.5B-Instruct (~1.5B parameters, ~3 GB in fp16). It is built for FIRST Tech Challenge (FTC) and FIRST Robotics Competition (FRC) teams who want a lightweight local coding helper for OpModes, autonomous routines, encoder/motor control, and OpenSCAD CAD scripts. Developed by FTC Team 24501 "ERROR" from Diyarbakır, Türkiye.

Türkçe özet — Mithra-2B, FTC ve FRC robotları için Java/C++ kodu, OpMode ve otonom taslakları, OpenSCAD/CAD betikleri yazan ve yazılım hatalarını çözmeye yardım eden küçük, açık kaynaklı bir kodlama modelidir. FTC Team 24501 "ERROR" (Diyarbakır) tarafından geliştirilmiştir.


✨ At a Glance

Model name Mithra-2B (the "2B" is a series name; actual size ≈ 1.5B parameters)
Type Causal (decoder-only) language model, instruction-tuned
Base model Qwen/Qwen2.5-Coder-1.5B-Instruct
Fine-tuning LoRA (r=16, α=32) with TRL SFTTrainer, merged into base weights
Languages Turkish 🇹🇷, English 🇬🇧
Code languages Java, C++, Python, JavaScript, HTML, CSS, OpenSCAD
Focus FTC SDK, FRC / WPILib style robot code, autonomous, CAD scripting
Size ~3 GB (fp16) — runs on a free Colab T4 or a modest local GPU
License MIT
Made by FTC Team 24501 "ERROR", Diyarbakır, Türkiye
Funded by TED Diyarbakır

🎯 What can Mithra-2B do?

🔧 Task 💬 Example prompt
FTC OpMode (Java) "FTC için encoder ile motoru hedef konuma götüren bir OpMode yaz."
Drive system "Write a mecanum TeleOp drive OpMode for FTC in Java."
Autonomous draft "Write an autonomous routine that drives forward 24 inches and stops."
Sensor logic "Read a color sensor and stop the robot when it sees yellow."
CAD / 3D printing "Write an OpenSCAD script for a motor mount with 4 screw holes."
Debugging "Bu Java hatasını açıkla ve düzelt: NullPointerException hardwareMap…"

🚀 Quick Start

pip install torch transformers accelerate
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

REPO = "Bedirhn77/Mithra-2B"

SYSTEM_PROMPT = (
    "Senin adın mithra. Türkiye, Diyarbakır merkezli FTC Team 24501 'ERROR' (ftcerror.com) "
    "tarafından geliştirilmiş bir yapay zeka kodlama ve robotik asistanısın. Görevin FTC ve FRC "
    "robotları için Java ve C++ kodları yazmak, CAD tasarımları için Python/OpenSCAD betikleri "
    "üretmek ve yazılım sorunlarını çözmektir. Kendinden bahsederken sadece mithra olduğunu ve "
    "ERROR takımı tarafından Diyarbakır'da oluşturulduğunu söylemelisin."
)

tok = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.float16, device_map="auto")

prompt = (
    f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
    "<|im_start|>user\nFTC için encoder ile motoru hedef konuma götüren bir OpMode yaz.<|im_end|>\n"
    "<|im_start|>assistant\n"
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs,
    max_new_tokens=400,
    do_sample=True,
    temperature=0.4,
    top_p=0.9,
    eos_token_id=tok.convert_tokens_to_ids("<|im_end|>"),
)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

💡 Tip: Use the system prompt above so the model keeps its mithra identity. Low temperature (0.2–0.4) gives more reliable code.


🧪 Training Details

Dataset First 1,200 rows of iamtarun/code_instructions_120k_alpaca
Format ChatML with the mithra identity system prompt
Method LoRA (PEFT): r=16, lora_alpha=32, lora_dropout=0.05, targets q_proj, v_proj
Trainer TRL SFTTrainer
Precision fp16 mixed precision
Epochs 2
Effective batch size 16 (2 × 8 gradient accumulation)
Learning rate 2e-4, cosine schedule
Max sequence length 768
Hardware Free Google Colab, NVIDIA T4 — about 20–40 minutes
After training LoRA merged into the base model with merge_and_unload

⚠️ Limitations & Responsible Use

  • Small model, small dataset. Only 1,200 training examples; it does not deeply cover the FTC SDK or WPILib and may invent functions that do not exist. Always check against the official docs.
  • Always test. Review generated code and run it in a simulator or a safe setup before using it on a real robot. Do not run unverified code in safety-critical situations.
  • Turkish quality can be weaker than English.
  • Identity (mithra / ERROR / Diyarbakır) comes from the system prompt; without it the model may answer as the base model.
  • Long, complex projects: consistency is limited compared to large models.
  • No formal benchmark has been run. The model was only checked manually with FTC/FRC sample prompts and identity questions.
  • Not designed for medical, legal or financial advice.

❓ FAQ

What is Mithra-2B? A small open-source coding and robotics language model for FTC and FRC teams, fine-tuned from Qwen2.5-Coder-1.5B-Instruct by FTC Team 24501 "ERROR".

Who made Mithra-2B? FTC Team 24501 "ERROR" from Diyarbakır, Türkiye (ftcerror.com), led by Bedirhan, with support from TED Diyarbakır.

Is it really 2 billion parameters? Yes. "2B" is a Parametre. The train size is about 0.5B parameters and the base model is Qwen2.5-Coder-1.5B-Instruct.

Can it write FTC OpMode code? Yes, it can draft Java OpModes (TeleOp, autonomous, encoder control). Review and test everything before running on a robot.

Can it make CAD models? It can write OpenSCAD and Python scripts for simple parts like mounts, brackets and plates.

Does it speak Turkish? Yes — Turkish and English, with English usually stronger.

Can I run it for free? Yes. It runs on a free Colab T4 GPU and many consumer GPUs (~3 GB in fp16).

Can I fine-tune it further on my team's code? Yes. It is MIT-licensed and intended as a starting point for team-specific fine-tuning with your own robot libraries.

Which license? MIT.


🔎 Keywords

FTC coding AI · FRC coding assistant · FIRST Tech Challenge LLM · FIRST Robotics Competition AI · FTC OpMode generator · FTC autonomous code · FTC SDK Java · WPILib helper · OpenSCAD AI · robot code generator · small coding LLM · Qwen2.5-Coder fine-tune · LoRA coding model · Turkish coding model · Türkçe kodlama modeli · FTC yapay zeka · robotik kod asistanı · FTC Team 24501 ERROR


📚 Citation

@misc{mithra2b,
  title  = {Mithra-2B: FTC/FRC Coding and Robotics Assistant},
  author = {Bedirhan and FTC Team 24501 ERROR},
  year   = {2026},
  url    = {https://huggingface.co/Bedirhn77/Mithra-2B}
}

🙏 Acknowledgements

📬 Contact

FTC Team 24501 "ERROR" — Diyarbakır, Türkiye · 🌐 ftcerror.com

⭐ If Mithra-2B helps your team, give the repo a like on Hugging Face! ⭐

Downloads last month
454
Safetensors
Model size
2B params
Tensor type
F16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 1 Ask for provider support

Model tree for Bedirhn77/Mithra-2B

Adapter
(237)
this model
Adapters
1 model

Dataset used to train Bedirhn77/Mithra-2B