Repository navigation
feat(007): trained inference path for seed-morph78 - #106
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e028323fb1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| raise InferModelError("family probabilities do not match families") | ||
| best = max(range(len(families)), key=lambda i: family_probs[i]) | ||
| vector_blob = ",".join(f"{v:.6f}" for v in vector).encode("utf-8") | ||
| unbind_ok = bool(atoms) and all(a["filler_pred"] == a["filler"] for a in atoms) |
There was a problem hiding this comment.
When the filler head predicts every token correctly but the role head predicts any wrong positional role, this expression still reports unbind_ok: true. Spec 007 defines unbinding as role-and-filler exactness, so downstream consumers can accept a structurally incorrect result; require both role_pred == role and filler_pred == filler for every atom.
Useful? React with 👍 / 👎.
| weight = model_dir / WEIGHT_FILE | ||
| if not weight.is_file(): | ||
| raise InferModelError(f"missing {weight}") |
There was a problem hiding this comment.
When safetensors is unavailable or saving it fails, save_heads intentionally produces heads.pt, and the checkpoint contract declares both formats valid. This hardcoded lookup therefore rejects a valid output from the repository's own trainer as missing; inference should select and load either artifact, as the existing trunk-forward evaluator does.
Useful? React with 👍 / 👎.
| offs = _offsets(tok, text) | ||
| atoms = [] | ||
| for i, token in enumerate(tokens): | ||
| idxs = pool_indices(states.size(0), atom_token_index(text, token, offs)) |
There was a problem hiding this comment.
For repeated tokens such as go go, or a token that first appears as a substring such as the a in cat a, atom_token_index ultimately uses text.find(token) and resolves the first matching substring every time. Multiple positional atoms consequently pool the same or an unrelated hidden state while being labeled as different pos_i roles, corrupting their role and filler predictions; use the character span of the enumerated whitespace occurrence instead.
Useful? React with 👍 / 👎.
| tokens = text.split() | ||
| with torch.no_grad(): | ||
| enc = tok([text], truncation=True, max_length=MAX_LEN, return_tensors="pt") |
There was a problem hiding this comment.
For inputs exceeding the tokenizer's MAX_LEN, tokens retains every whitespace token while the encoder truncates the sequence. Atoms beyond the retained offset mapping then get no matching indices and pool_indices silently substitutes token state 1, so the packet fabricates predictions for text the model never encoded and can even report a false successful unbind; reject, omit, or explicitly mark those truncated atoms unsuccessful.
Useful? React with 👍 / 👎.
1 participant
Same change as #105, rebuilt on
mainafter #104 was squash-merged. #105 had adjacent-insertion conflicts in CHANGELOG, ROADMAP and similar files that the API can't resolve with a merge commit. The file tree is byte-identical to the tested commit.See #105 for the full description, Spark receipt, findings and validation (747 passed; mkdocs strict PASS). Supersedes #105.