8000 Corpus: OpenHermes held out on its own; self-knowledge check finds its files by gt50 · Pull Request #5 · gt50/mini-AGI · GitHub
[go: up one dir, main page]

Skip to content

Corpus: OpenHermes held out on its own; self-knowledge check finds its files - #5

Merged
gt50 merged 1 commit into
mainfrom
corpus-lanes
Oct 4, 2026
Merged

gt50 merged 1 commit into
mainfrom
corpus-lanes

Conversation

@gt50
@gt50 gt50 commented Oct 4, 2026
Copy link
Copy Markdown
Owner

Summary

Two path bugs in the chat lanes of python -m corpora all, found while building the full corpus on Windows.

  • OpenHermes' held-out set was deleted on every full build. OpenHermes held out into data/val/chat. The generated chat lane (primal chat plus the self-knowledge turns) runs afterwards with expand --only chat, which clears the .txt files in data/{train,val}/chat, and that took OpenHermes' 400 held-out files with it.
    • OpenHermes now holds out into data/val/chat_hermes, the separate held-out subject the README's results already report (chat_hermes 0.924).
    • Its training data stays in data/train/chat/hermes, inside the chat subject as before.
  • The self-knowledge lane always "produced nothing". It was checked at data/train/self-knowledge, a folder nothing writes, so every full build ended with lanes that produced nothing: self-knowledge and exit code 1, even though the lane had been written to data/train/chat.
    • It's now checked with a glob, data/train/chat/chat-*.txt. Checking the folder itself wouldn't work: it also holds hermes/, so it would read as built before the lane ever was.
    • _files() replaces the directory walk in both the skip check and the summary.
  • serve.py's priming read the same missing folder, so it always served with no priming. It reads data/train/chat now.

The first passage it picks is a code-explanation exchange rather than a question about the model itself. The generated lane mixes both kinds and doesn't name them apart the way upstream's self-* files did. That's left as is.

Test plan

  • Every lane check, after a full build: all eight see their files (self-knowledge: 2,180 chat-*.txt)
  • python -m corpora all --only chat --force writes 400 held-out files to data/val/chat_hermes, and data/val/chat keeps its 22
  • data/val now has the eight subjects the README reports: arithmetic, chat, chat_hermes, chess, code, reasoning, stories, wikipedia
  • serve.load_prime(1024) returns a 566-character balanced passage, where before it returned nothing
  • pytest: 22/22
  • CI green

🤖 Generated with Claude Code

…s files

`python -m corpora all` had two path bugs in the chat lanes:

- OpenHermes wrote its held-out set into data/val/chat, and the generated
  chat lane (primal chat + self-knowledge) runs `expand --only chat`
  afterwards, which clears the .txt files in data/val/chat - silently
  deleting OpenHermes' 400 held-out files on every full build. It now
  holds out into data/val/chat_hermes, the separate held-out subject the
  README's results already report. Its training data stays in
  data/train/chat/hermes.
- The self-knowledge lane was checked at data/train/self-knowledge, a
  folder nothing writes, so every build reported "self-knowledge produced
  nothing" and exited 1 although the lane had been written to
  data/train/chat. It is checked by glob, data/train/chat/chat-*.txt -
  not by the folder, which holds hermes/ too and would read as built
  before the lane ever was.

serve.py's priming read the same missing folder, so it served unprimed;
it reads data/train/chat now.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@gt50
gt50 merged commit 0fa8a9f into main Oct 4, 2026
2 checks passed
@gt50
gt50 deleted the corpus-lanes branch October 4, 2026 22:48
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.

1 participant

0