agentleFS
Sign inStart free
zjunlpzjunlp/Mechanist/skills/experiment-tips/finetune-hyperparameter-sweep/SKILL.md

finetune-hyperparameter-sweep

Skill75 starsChanged 4 months agoMITView on GitHub

Fine-tuning LR protocol — full FT, LoRA / QLoRA / DoRA / PEFT adapter, across SFT, DPO, and GRPO / PPO / RL objectives. Fires on ANY fine-tune, including (especially) when the plan already fixes a learning rate or copies one from a reference paper: a fixed LR is one grid point, never the answer, so LR is always swept over a wide range. The only test of an LR is whether the run hits the pass criteria in `task.md` or the milestone''s own claim criteria — loss, grad norm, and reward curves are never acceptance evidence, no matter how good they look. Criteria not met = change LR and re-run. Triggers: `learning_rate`, `SFTTrainer`, `DPOTrainer`, `GRPOTrainer`, `LoraConfig`, `peft_config`, `full fine-tune`, `SFT`, `DPO`, `GRPO`, `RLHF`, `QLoRA`, `DoRA`, `PEFT`, `adapter`.

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Criteria not met → change LR and re-run: Keep going until an LR passes.

    line 12

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Fine-Tuning LR Sweep

  1. Any fine-tune → always sweep LR over a wide range. Whether the plan hard-codes an LR, copies one from a paper, or leaves it open makes no difference: that value is one grid point, never the answer.
  2. The only test of an LR is the criteria — the pass criteria in task.md, or the claim criteria of the relevant experiment-plan milestone. Nothing else counts as acceptance evidence: a converged loss, a healthy grad norm, a rising reward, a "reasonable-looking" curve certify nothing. Criteria not met = LR not acceptable, no matter how good the training curves look.
  3. Criteria not met → change LR and re-run. Keep going until an LR passes. Only after the whole LR range has failed should you touch anything else.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free