Skip to content

Fix: score final partial window in sliding window eval #124

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
0hq merged 1 commit into from
Mar 19, 2026
Merged

Fix: score final partial window in sliding window eval #124

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions records/track_10min_16mb/2026-03-19_SlidingWindowEval/train_gpt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -856,9 +856,9 @@ def eval_val_sliding(
seq_len = args.train_seq_len
total_tokens = val_tokens.numel() - 1

# Build windows; skip any too short to score a full stride
# Build windows; include final partial window if it has at least 1 token
window_starts = [ws for ws in range(0, total_tokens, stride)
if min(ws + seq_len, total_tokens) - ws >= stride]
if min(ws + seq_len, total_tokens) - ws >= 1]
total_windows = len(window_starts)

# Distribute across ranks
Expand Down Expand Up @@ -899,7 +899,7 @@ def eval_val_sliding(

for i, ws in enumerate(batch_ws):
wlen = wlens[i]
s = 0 if ws == 0 else wlen - stride
s = 0 if ws == 0 else max(wlen - stride, 0)
scored_nll = nll[i, s:wlen].to(torch.float64)
loss_sum += scored_nll.sum()
token_count += float(wlen - s)
Expand Down