Tiny Tool Use is a minimal, open-source library for LLMs to make reliable, auditable tool calls. Supports SFT, DPO, and synthetic data — all driven by simple JSON config. Fast setup, strong evals, and ready for real-world prototyping.
👋 Hey Product Hunt!
At Bagel Labs, we believe the future of advanced AI systems depends on their ability to reason with external tools, APIs, and data sources. But when we tried building with existing “tool-use” stacks, we ran into the same issues every time: brittle code, no reproducibility, and zero shared benchmarks.
So we built Tiny Tool Use — a minimal, MIT-licensed open-source library that turns adapting LLMs for robust, auditable tool calls into a config-only workflow. No fragile scaffolding. No hidden state. Just one JSON file and a single CLI.
🚀 Why we built it
Adapt open source LLMs in minutes, not days — from config to full fine-tune
Make tool calls auditable via explicit schemas and benchmarked outputs
Scale synthetic data generation via Teacher Mode — models generate, remix, and train on their own traces. Tool use is what makes models like o3 so great.
🛠️ What makes it different
• Configuration-only pipelines — models, tools, datasets, hyper-params all live in JSON
• One-flag training mode swap — "method": "sft" | "dpo" | "teacher_mode"
• Synthetic data at scale — blend real + generated traces via "real_fraction"
• Full evaluation suite — TensorBoard, Berkeley Function Calling Leaderboard (BFCL), and detailed metrics
• Production-ready — LoRA adapters, GPU-agnostic trainer, model merge utilities
🔍 Evaluations that matter
Every training run includes:
• Tool selection accuracy
• Format correctness
• Execution success
• Response quality
And every run is logged with full provenance: config, seed, Git commit.
🏗️ What we’re proud of
<400 lines of core logic — readable, hackable, extensible
Full LoRA support from 0.6B to 70B+
Community contributions already landing (new tool schemas, eval reports)
Seamless end-to-end: train, merge, benchmark, share
Whether you're a researcher, builder, or just LLM-curious — we built Tiny Tool Use so anyone can take models from static text prediction to dynamic tool reasoning.
We’re excited to launch today and can’t wait to see what you build with it. Check out the repo here!
👉 GitHub: https://github.com/bagel-org/bag...
@kyrannio This is exactly what the ecosystem needed! The config-only approach is brilliant, it removes so much boilerplate complexity. Quick question, how does the Teacher Mode synthetic data generation compare to traditional fine-tuning in terms of model performance? Are you seeing better generalization?
Report
Can’t wait to see it all in action so I think I will start reading all the documentation
Coverr
Femme Stock
@kyrannio Congratulations Kiri! So happy to get the early access.
Coverr
@zeng Thanks, Zeng! We are so thrilled to have you!
Smoopit
@kyrannio This is exactly what the ecosystem needed! The config-only approach is brilliant, it removes so much boilerplate complexity. Quick question, how does the Teacher Mode synthetic data generation compare to traditional fine-tuning in terms of model performance? Are you seeing better generalization?
Can’t wait to see it all in action so I think I will start reading all the documentation
Coverr
@unity_eagle woohoo! Sounds great, Unity! We are excited to see what you build :). Thanks for joining us!
This launch is a dream come true for developers
Coverr
@rufus87078959 Thanks Rufus. We can’t wait to see what you build! :)
Clean utility 🔧🧠 Simplifying tool use with open-source LLMs hits the sweet spot for devs and tinkerers.
Can't wait to start working with Bagel Labs tools and AI!
This is the future of AI systems!