HRM-Text 101 Tutorial

追加: 2026-05-26

[00:00:08]Hey everyone, this is Yasin from Sapion Intelligence. Welcome to HRM text 101.

[00:00:14]HRM text is a 1 billion parameter hierarchal reasoning language model. The text variant of HRM built on the same two time scale recurrence that made HRM effective on symbolic reasoning. In this video, I will walk you through downloading the base checkpoint, fine-tuning it on a real task, and evaluating the results end to end on a single GPU. So, let's get started. The code lives on the this GitHub repository.

[00:00:41]Uh, everything you need, the training script, the HRM backbone, data tooling, and a fully worked demo config.

[00:00:53]The Redmi has the full quick start with environmental setup. I assume your CUDA and PyTorch are ready and skip straight to the interesting parts. Download the checkpoint, prepare the data, fine-tune and evaluate. The pre-trained HRM text base model leaves on hugging face. One command pulls it down.

[00:01:27]The checkpoint is in FSTP2 format. You will see FSTP2 epoch directory with the shorted weights and a couple of config files. We'll point pre-train.py at this in the next step. Now let's prepare the data. You'll find tune on spider a text tossql benchmark with a thousand depth questions across 200 databases.

[00:01:55]This script does three things. First, it converts the spider training split into few shot JSONL. Each example carries three in context demos from other databases. So, the model learns to ground SQL in whatever schema you hand it.

[00:02:11]Second, it tokenizes the JSON with our EP tokenizer about 8.7 million training tokens across 7,000 examples. Third, it packs the tokens into a box. This packing is what lets us hit high GPU utilization with variable length sequences.

[00:02:30]And that's it. Data leaves in shared memory ready for training. So now time to finetune one command.

[00:02:47]The base checkpoint loads and training begins. We are doing full supervised finetuning on a single H00 global batch size of 4096 5 epox and 7,000 examples that takes around 15 minutes.

[00:03:05]Done. The final checkpoint is in / CKPTs/ SFT/ So let's see what we got.

[00:03:12]The AIS script runs the fine tuned model on spider defet and scores execution accuracy. Does the generated skill return the right answer when you actually run it against the database?

[00:03:32]Across the full def set, the finetune model hits around 62% execution accuracy, up from around 8% for the base checkpoint. So what does it mean in practice? Before finetuning, the base model sometimes doesn't even return SQL.

[00:03:48]Here is an answer with the number 254 to what is the total number of singers on harder queries? it falls into schema token loops or outputs JSON L JSON like noise where SQL should be. Sometimes the failure is subtle unnecessary joins that dilute the count or the classic hallucinated column name. After fine-tuning all five of these are correct. You get a model that has actually learned to read schemas and grounding in them. And that's it. You have gone from a fresh checkpoint to a fine-tuned hierarchal reasoning model.

[00:04:26]If you want to follow what we are working on or share what we have built with HRM, come find us. Links are all in the description. See you there.