Codex Now Automates End-to-End Machine Learning Experiments

Codex is Open Sourcing AI models

OpenAI’s Codex, integrated with the Hugging Face Skills repository, now automates end-to-end machine learning experiments, streamlining the process from data preparation to model deployment. This integration enables Codex to handle tasks like fine-tuning language models, monitoring training metrics, evaluating checkpoints, and even converting models to GGUF for local use.

Why This Matters

Currently, ML experimentation requires significant manual intervention, from configuring hardware and writing training scripts to monitoring progress and debugging failures. A single failed training run can cost hundreds of dollars in GPU time and weeks of engineering effort. Automating these processes reduces costs and accelerates the development cycle, making advanced ML techniques more accessible.

Key Insights

• Codex leverages AGENTS.md files: Unlike Claude Code which uses ‘Skills’, Codex utilizes AGENTS.md files to define specialized tasks.
• SFT, DPO, and RLHF support: The system supports supervised fine-tuning, direct preference optimization, and reinforcement learning with verifiable rewards.
• Hugging Face Integration: Codex seamlessly integrates with Hugging Face tools like Jobs and Trackio for training and monitoring, and supports model publishing to the Hub.

Working Example

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("burtenshaw/qwen3-codeforces-cots-sft")
tokenizer = AutoTokenizer.from_pretrained("burtenshaw/qwen3-codeforces-cots-sft")

Practical Applications

• Research Labs: Automate hyperparameter sweeps and model evaluations, allowing researchers to focus on higher-level experimentation.
• Pitfall: Over-reliance on automated systems without understanding the underlying configurations can lead to suboptimal model performance or unexpected behavior.

References:

• https://huggingface.co/blog/hf-skills-training-codex