slop.boutique

A lab for computation and intelligence as material and medium. Run by Manuel and Ramon.

029 · 23 October 2026 · machine-learning · agents · education

Kiln Studio

An IDE for LLM post-training programs: write SFT, DPO and RL pipelines in JavaScript, see them as a graph, catch leaks, watch a simulated run, and export code for Unsloth, TRL, prime-rl, verl or PyTorch.

Previewin your browser
A GRPO training program on GSM8K next to its simulated run: reward, policy entropy, zero-variance groups and learning-rate curves over 400 steps, and the utilisation of 8 learner and 4 rollout GPUs

In early access for subscribers until 22 November 2026. Subscribe to run it now.

Subscribe →

An IDE for LLM post-training programs. Write a training run in a small JavaScript language (kiln for training, quarry for data), and the studio traces it into a graph, compiles it, checks it for leaks and mistakes, and runs a simulated version so you can watch it train. It is a prototype, separate from Kiln Workbench: that one is a compiler workbench for GPU kernels, this one is about SFT, DPO and RL pipelines.

  • Programs as graphs: SFT, DPO, GRPO and DAPO objectives, rewards, graders, LLM judges, evalsets, splits, stages, sweeps and gates become a colour-coded graph, as written or as the lowered IR, with dev, test and sealed rows drawn as their own edges.
  • Problems before you run: it flags test rows that reach training data, early stopping on a test metric, rewards that throw, and inserts decontamination against your evalsets.
  • Plan: memory per GPU, compute and wall time on a chosen backend, and the compiler passes that ran (fusion, activation checkpointing, FSDP sharding, communication overlap).
  • Watch it train, then break it: a simulated run at 1×, 4× or 16× with reward, entropy, loss and throughput curves, rollout groups scored by your graders, per-GPU utilisation, eval results with intervals and gate decisions. Inject out-of-memory, malformed outputs, entropy collapse or preemption and see what happens.
  • Data recipes: generate problems with a teacher model, verify, deduplicate and decontaminate them, and browse the rows at every step.
  • Export: generated, readable code for Unsloth, TRL, prime-rl, verl and native PyTorch, with a lock file, a runbook and your lambdas bundled for the runtime.

Twenty example programs come with it, from SFT in one expression and a program with three mistakes to agentic GRPO with tools on a MoE, distilling traces with best-of-n, a rejection-sampling fine-tune loop and one program on five backends, plus a library of presets for training, data, evals, gates and compute targets.

Training runs are simulated in the browser; graders run on synthetic completions. The interface uses Spline Sans and Spline Sans Mono from Google Fonts (SIL Open Font License), React, htm and Acorn.

The full studio: the program 'Splits, dev evals and a regression gate' in the editor, its traced graph of 21 colour-coded nodes on the right, and two decontamination notes in the Problems panel
The full studio: the program 'Splits, dev evals and a regression gate' in the editor, its traced graph of 21 colour-coded nodes on the right, and two decontamination notes in the Problems panel
Close-up of a finished simulated GRPO run: eight charts (reward by grader, policy entropy, clipped ratio, response length, zero-variance groups, learning rate, gradient norm, learner throughput), rollout groups scored by the graders, and per-GPU memory and utilisation
Close-up of a finished simulated GRPO run: eight charts (reward by grader, policy entropy, clipped ratio, response length, zero-variance groups, learning rate, gradient norm, learner throughput), rollout groups scored by the graders, and per-GPU memory and utilisation
Export tab for 'One program, five backends': a backend matrix (Unsloth, TRL, prime-rl, verl, native PyTorch), the generated file tree, and the generated Unsloth train.py with comments that point back to the program's lines
Export tab for 'One program, five backends': a backend matrix (Unsloth, TRL, prime-rl, verl, native PyTorch), the generated file tree, and the generated Unsloth train.py with comments that point back to the program's lines
Graph of 'A program with three mistakes': test rows flowing into the training data are drawn as red leak edges from split.test through data.mix to train
Graph of 'A program with three mistakes': test rows flowing into the training data are drawn as red leak edges from split.test through data.mix to train
Plan tab: memory per GPU, compute and time estimates for the SFT stage, and the compiler passes that ran (decontamination, fused linear and log-prob, activation checkpointing, FSDP sharding, communication overlap)
Plan tab: memory per GPU, compute and time estimates for the SFT stage, and the compiler passes that ran (decontamination, fused linear and log-prob, activation checkpointing, FSDP sharding, communication overlap)
An injected out-of-memory failure: the SFT stage stops at step 210 of 600 with an explanation, loss and learning-rate curves up to that step, the GPUs, and the event log
An injected out-of-memory failure: the SFT stage stops at step 210 of 600 with an explanation, loss and learning-rate curves up to that step, the GPUs, and the event log
Evals tab after a run: the dev and test evalsets, resolved and exact scores at baseline and step 300 with intervals, results by customer tier, and the regression gate that passed
Evals tab after a run: the dev and test evalsets, resolved and exact scores at baseline and step 300 with intervals, results by customer tier, and the regression gate that passed
The Open dialog with the example programs grouped by goal: Start here, Evaluation, Reinforcement learning, Preference tuning, Export
The Open dialog with the example programs grouped by goal: Start here, Evaluation, Reinforcement learning, Preference tuning, Export
Data tab for 'Synthetic problems, verified': the chain of generate, flatten, dedup, verify, filter, semantic dedup and decontamination steps with row counts, and a page of the generated problems
Data tab for 'Synthetic problems, verified': the chain of generate, flatten, dedup, verify, filter, semantic dedup and decontamination steps with row counts, and a page of the generated problems
The preset library: training recipes, data recipes, evals and gates, with the GRPO on math preset's parameters, usage and source
The preset library: training recipes, data recipes, evals and gates, with the GRPO on math preset's parameters, usage and source
Graph of 'Pipeline: SFT, DPO sweep, GRPO': three stages chained through a pick and a gate
Graph of 'Pipeline: SFT, DPO sweep, GRPO': three stages chained through a pick and a gate