klarsklarsThe model-first research environment
Open any model. Keep every reason.
Open a model from where you found it, on GitHub or Hugging Face, and land in a working representation of it, in your browser or on your desktop. Run it on the compute it needs, and keep the reason behind every checkpoint: what it came from, what else was tried and why it was chosen.
The repository path github.com/your-lab/model opens as a small model graph: embed, a block repeated 12 times, ln_f and head, shown at checkpoint 12. Beneath it, the model's checkpoint lineage runs over 14 thousand training steps. Four sibling runs branch off, each with its reason. lr 3e-4: diverged at step 3.2k. Plus warmup 2k: stable, eval 2.98. Data mix v2: eval regressed. d_ff 4096: no gain over ckpt-07. One lit path runs through ckpt-03 and ckpt-07 to the chosen checkpoint, ckpt-12. Its record reads: from ckpt-07 with rmsnorm; tried lr, warmup, data and d_ff; why, lowest eval loss, 2.91; after, ckpt-14, an sft run.
01The thinking
A model is a graph, not a text file.
Today one model lives in five places: its source, a print(model) dump, a tracker tab, a checkpoints folder and an SSH session. None of them knows about the others. klars puts the model first and organizes everything else around it. Code stays the source of truth, opened when you need it.
- 01 · Model first
Work in model space.
The model is the primary layer: its graph, its source, its runs and the machine it runs on, joined on every node. Code is the second modality.
- 02 · Any model, any compute
Open it where you found it.
Start from a GitHub repo or a Hugging Face model. Run it in a sandbox with everything installed, where GPUs attach and detach as resources: pay for GPU time, not while you code.
- 03 · Every reason kept
Checkpoints with a history.
Every checkpoint keeps where it came from, the sibling runs that were tried and why they stopped, and why this one was chosen.
- 04 · A research harness
Agents that improve your model.
A harness that pursues an objective you set: propose, run, evaluate, compare, remember and loop, within budgets and review gates, with you steering.
02Built so far
It starts with tensorViz, the visual layer for machine learning.
tensorViz draws a PyTorch model as a live graph tied to its code, turns graph edits into code diffs, and gives coding agents the model instead of the files. It is available today in VS Code, with an agent plugin in preview.