Skip to content

Fix ZeroDivisionError and NaN propagation in reward model metrics #3779

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

Open
Mr-Neutr0n wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix ZeroDivisionError and NaN propagation in reward model metrics #3779

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
31 changes: 27 additions & 4 deletions model/model_training/metrics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,15 @@ def reward_accuracy(eval_pred):
logits_batch = b_logits[b_labels == i]
pos_scores.append(logits_batch[0])
neg_scores.append(logits_batch[-1])

if len(pos_scores) == 0:
return {
"pos_score": 0.0,
"neg_score": 0.0,
"score_diff": 0.0,
"accuracy": 0.0,
}

pos_scores = np.array(pos_scores).reshape(-1, 1)
neg_scores = np.array(neg_scores).reshape(-1, 1)

Expand Down Expand Up @@ -43,10 +52,17 @@ def kendall_tau(eval_pred):
# b_logits = b_logits[:truncated_logits]
for i in np.unique(b_labels):
logits_batch = b_logits[b_labels == i]
if logits_batch.size < 2:
continue
pred_rank = np.argsort(logits_batch)
true_rank = np.arange(logits_batch.size - 1, -1, -1)
tau += st.kendalltau(pred_rank, true_rank)[0]
bsize += np.unique(b_labels).size
result = st.kendalltau(pred_rank, true_rank)[0]
if not np.isnan(result):
tau += result
bsize += 1

if bsize == 0:
return {"kendalltau": 0.0}

return {"kendalltau": tau / bsize}

Expand All @@ -61,10 +77,17 @@ def spearmanr(eval_pred):
b_logits = b_logits[b_logits != -100]
for i in np.unique(b_labels):
logits_batch = b_logits[b_labels == i]
if logits_batch.size < 2:
continue
pred_rank = np.argsort(logits_batch)
true_rank = np.arange(logits_batch.size - 1, -1, -1)
score += st.spearmanr(pred_rank, true_rank).statistic
bsize += np.unique(b_labels).size
result = st.spearmanr(pred_rank, true_rank).statistic
if not np.isnan(result):
score += result
bsize += 1

if bsize == 0:
return {"spearmanr": 0.0}

return {"spearmanr": score / bsize}

Expand Down