Skip to content

Adjust inference reward scales #703

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
merged 1 commit into from
May 1, 2025
Merged

Adjust inference reward scales #703

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
11 changes: 5 additions & 6 deletions prompting/rewards/exact_match.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
MAX_VERIFY_TOKENS = 30
NO_EOS_PENALTY = -0.1
INCORRECT_PENALTY = -2
NOT_ENOUGH_TOKENS_PENALTY_SCALE = 0.1
MIN_SMOOTH_PENALTY_SCALE = 0.6
MIN_TIME_PENALTY_SCALE = 0.3
VERIFICATION_THRESH_CONTAINS = 0.96
Expand Down Expand Up @@ -67,6 +68,7 @@ async def reward( # noqa: C901
# Iterate over each miner response.
for chunks, timings, chunk_dicts_raw, uid in zip(all_chunks, all_timings, all_chunk_dicts_raw, uids):
penalty = INCORRECT_PENALTY
reward_scale = 1.0
try:
# If no response is provided, apply full penalty.
if not chunks:
Expand All @@ -82,10 +84,8 @@ async def reward( # noqa: C901
continue

if completion_length < MIN_VERIFY_TOKENS:
# Not enough tokens to verify, set reward to 0.
rewards.append(0)
timing_verified.append([-1.0])
continue
# Not enough tokens to verify, still proceed to verification with scaled reward if checks will pass.
reward_scale = NOT_ENOUGH_TOKENS_PENALTY_SCALE

eos_idx = completion_length
verify_indices = self.sample_verification_indices(completion_length)
Expand Down Expand Up @@ -142,8 +142,7 @@ async def reward( # noqa: C901
# Min-max scale logits reward, e.g from [0.95; 1.0] to [0.0, 1.0].
score_sim_mean = self.rescale(score_sim_mean, min_value=VERIFICATION_THRESH_SIM)
score_contains_mean = self.rescale(score_contains_mean, min_value=VERIFICATION_THRESH_CONTAINS)
logits_score = (score_sim_mean + score_contains_mean) / 2
rewards.append(logits_score * smooth_reward)
rewards.append(score_sim_mean * score_contains_mean * smooth_reward * reward_scale)
except BaseException as e:
logger.debug(f"Miner {uid} failed to pass logits check: {e}")
rewards.append(penalty)
Expand Down