Hugging Face smol course: Review

I’m Sofia, a CS student with no budget for rented GPU heroics. That is why Hugging Face’s smol course caught me: post-training lessons built around small models that can run on consumer hardware, with no paid service required. The published material is useful. The course is also unfinished. You need both facts before planning around it.

Four units are live: instruction tuning with SFT and chat templates, evaluation with standard and domain-specific benchmarks, preference alignment through DPO and related methods, and vision-language models. That sequence fills a gap most free catalogs leave open. Plenty of courses explain what an LLM does; few walk through changing one after pretraining and measuring whether the change helped.

The hardware promise is the best part. SmolLM3 and SmolVLM2 keep the exercises within reach of ordinary machines, and the repository says paid services are unnecessary. The project is open source, community-built, and had 6.7k stars with active commits at our August 2026 check. Discord study groups, a public model leaderboard, and pull-request contributions make it feel more like a workshop than a passive video playlist.

Now the warning in large type: Units 5, 6, and 7 are missing. The syllabus labels reinforcement learning for October, synthetic data for November, and an award unit for December. Those dates referred to 2025. None was live by August 2026, and no replacement timetable appeared. The full Certificate of Completion requires every unit plus a final project, so it is effectively unavailable. A free Fundamentals certificate after Unit 1 is reachable; the larger promise is not.

There are smaller rough edges. The v1 and v2 tables of contents coexist, which makes the site look less settled than the active repository suggests. Instructor names are not published. The prerequisites are also real: Python, ML basics, PyTorch experience, and a working understanding of transformer architecture. Hugging Face routes underprepared learners to its LLM Course first, and I agree. Prompt engineering is not the topic here either.

Independent written reviews are thin, so I would not pretend there is a deep learner consensus. The available catalog checks praise the same things I do: rare systematic coverage of post-training, free access, and realistic hardware. They also point to the unfinished syllabus and impossible completion certificate. That is why a 3.8 feels fair rather than harsh.

Use the LLM Course as the broader prerequisite. DeepLearning.AI’s Post-training of LLMs is a shorter video alternative, although its labs now sit behind Pro. Stanford CS224n goes deeper on RLHF and DPO theory. None matches this course’s specific promise of doing the work locally with small models.

For me, that local-first design is the reason to stay despite the gaps: I can repeat an experiment without watching a cloud-credit meter tick upward.

My recommendation: enroll for the four units that exist. They are free, hands-on, and unusually practical. Do not enroll for the missing RL unit or the full certificate. Check the official syllabus first; if Units 5 and 6 finally shipped, the score and recommendation deserve another look.