Own the weights, own your roadmap
Stop worrying about API deprecations, sudden price hikes, or stealth prompt changes. The trained adapter is a file on your disk.
SFT and reinforcement learning
Train open models on your actual work. Keep the weights.
Prompting only gets you so far. g factor takes an open model, teaches it with your best human examples (SFT), and refines it through trial-and-error in real workflows (GRPO). When training finishes, you own the weights—not a rented subscription.
SFT + GRPO
500+ open models
Portable LoRA weights
Your GPUs, your network
Why g factor
Why own the weights instead of renting an API?
Learn from real outcomes, not text vibes
Hook the model up to real compilers, simulators, and test suites. It gets rewarded when the output actually works, not when it sounds convincing.
A complete audit trail
Every dataset hash, GPU seed, checkpoint, and benchmark score is linked together. If something ever breaks in production, you can reproduce the exact run.
Capabilities
A governed path from expert data to an owned model skill.
Supervised fine-tuning
Distill trusted examples into a compact, portable adapter before reinforcement learning begins.
- Import JSONL or Hugging Face datasets
- Train LoRA adapters across open-model families
- Version model, tokenizer, data, and training configuration
GRPO in executable workflows
Improve behavior inside versioned environments that return task-specific rewards and terminal evidence.
- Typed multi-step actions through OpenEnv
- Native task, safety, quality, and efficiency signals
- Held-out tasks separated from training feedback
Distributed training with controls
Scale beyond a single machine while retaining resolved configuration, cost limits, and exact run provenance.
- Managed GPU topologies
- Explicit capacity and budget planning
- Portable checkpoints ready for independent qualification
Before you start
Common questions
What is LLM post-training?
It means taking a general-purpose open model (like Qwen, DeepSeek, or Llama) and teaching it to do a specific job exceptionally well. First, it learns from good human examples (SFT). Then, it practices in a simulator to master tricky edge cases (GRPO). You keep the resulting weights file.
When should I use SFT or GRPO?
Use SFT (imitation) when you have good logs, tickets, or examples of humans doing the job right. Use GRPO (reinforcement learning) when you have an automated test or environment that can grade the attempt. Most production skills use both: SFT teaches the basics, GRPO builds resilience.
Can I export and privately serve my fine-tuned model?
Yes. The LoRA adapter and merged checkpoints are standard open formats. You can take them, export them, and host them on your own private cloud or on-premises servers using vLLM or any standard engine.
Private beta
Turn your workflow into a skill your model can own.
Bring a model and a task that matters to your company. We'll tell you honestly if post-training fits.
Request Demo