Skip to content

Introduce Component-System for component optimization, add a dashboard to monitor activities#142

Open
LaurenceLong wants to merge 20 commits intokarpathy:masterfrom
LaurenceLong:master
Open

Introduce Component-System for component optimization, add a dashboard to monitor activities#142
LaurenceLong wants to merge 20 commits intokarpathy:masterfrom
LaurenceLong:master

Conversation

@LaurenceLong
Copy link

@LaurenceLong LaurenceLong commented Mar 10, 2026

Introduce Component-System for component optimization:

  • Add a dashboard to monitor activities
  • Provide UI operations make it easier for user to use
  • Support claudecode/codex/opencode, tested with opencode codeagent
  1. overview
image
  1. detail
image
  1. running logs
image

Copy link
Collaborator

@svlandeg svlandeg left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The dashboard looks awesome, but this PR with its 43 files changed is probably not very compatible with this repo's readme statement

The repo is deliberately kept small and only really has a three files that matter

😉

@LaurenceLong LaurenceLong force-pushed the master branch 2 times, most recently from 353d276 to 5a9c76f Compare March 11, 2026 02:11
@LaurenceLong
Copy link
Author

The dashboard looks awesome, but this PR with its 43 files changed is probably not very compatible with this repo's readme statement

The repo is deliberately kept small and only really has a three files that matter

😉

I agree. It doesn’t really fit the repo, so I won’t merge it, but I’d like to keep it around for anyone who wants to use it.

AutoResearch Agent and others added 18 commits March 12, 2026 13:37
Improve NaN loss handling and add beginner's guide resource
- Add max_grad_norm parameter (default 1.0) to TrainingSettings
- Import torch.nn.utils for clip_grad_norm_
- Apply gradient clipping before optimizer.step()
- This should improve training stability and potentially achieve better convergence

Target component: trainer
Expected benefit: Better training stability leading to improved val_bpb
Training:
- Make gradient clipping optional (max_grad_norm: float | None = None)
- Disabled by default to test training stability without clipping
- Can be enabled by setting max_grad_norm to a positive value

Workflow:
- Add _ralph_try_restore_worktree helper for consistent worktree reset logic
- Fix baseline waiting: allow planning on non-baseline branches without baseline
- Restore worktree on run failures and before P queueing when Ralph loop enabled
- Minor code formatting improvements in workflow.py
SwiGLU (SiLU Gated Linear Unit) is a proven improvement over ReLU²
from 'GLU Variants Improve Transformer' paper. Uses gated mechanism
with SiLU activation for better model capacity.

Changes:
- Replace c_fc/c_proj with gate_proj/up_proj/down_proj
- Use F.silu(up) * gate instead of ReLU²
- Adjust hidden_dim to maintain comparable parameters
- Update weight initialization for new layers
…st sync, track former_val_bpb, improve workflow merge order
…and concurrency handling

- Update PDCA documentation for clarity and consistency
- Add seed lifecycle and concurrency review documentation
- Track former_val_bpb for better signal evaluation
- Implement best_val_bpb tracking with history
- Add metrics recovery DCA for missing metrics
- Improve timeout handling (900s for DCA runs)
- Update web dashboard UI and templates
- Fix sync resolution and merge conflict handling
- Update protocol.md with simplified instructions
These checkpoint files are auto-generated by Jupyter and are already
listed in .gitignore. Removing them to keep the repository clean.
- Add prompt_audit directory to .gitignore
- Enhance worktree detection in run.py with separate workflows for worktree vs root modes
- Add merge resolution logic and temp branch handling in workflow.py
- Improve DCA stage with better baseline promotion tracking
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants