-
Notifications
You must be signed in to change notification settings - Fork 76
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.
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
Implements learned filter cascade #23
Merged
Merged
Changes from all commits
Commits
Show all changes
3 commits
Select commit
Hold shift + click to select a range
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,110 @@ | ||
| import numpy as np | ||
|
|
||
| import lotus | ||
|
|
||
|
|
||
| def importance_sampling( | ||
| proxy_scores: list[float], | ||
| sample_percentage: float, | ||
| ) -> tuple[list[int], list[float]]: | ||
| """Uses importance sampling and returns the list of indices from which to learn cascade thresholds.""" | ||
|
|
||
| w = np.sqrt(proxy_scores) | ||
| w = 0.5 * w / np.sum(w) + 0.5 * np.ones((len(proxy_scores))) / len(proxy_scores) | ||
| indices = np.arange(len(proxy_scores)) | ||
| sample_size = (int) (sample_percentage * len(proxy_scores)) | ||
| sample_indices = np.random.choice(indices, sample_size, p=w) | ||
| correction_factors = (1/len(proxy_scores)) / w | ||
|
|
||
| return sample_indices, correction_factors | ||
|
|
||
| def calibrate_llm_logprobs(true_probs: list[float]) -> list[float]: | ||
| """Transforms true probabilities to calibrate LLM proxies.""" | ||
| num_quantiles = 50 | ||
| quantile_values = np.percentile(true_probs, np.linspace(0, 100, num_quantiles + 1)) | ||
| true_probs = ((np.digitize(true_probs, quantile_values) - 1) / num_quantiles) | ||
| true_probs = np.clip(true_probs, 0, 1) | ||
| return true_probs | ||
|
|
||
| def learn_cascade_thresholds( | ||
| proxy_scores: list[float], | ||
| oracle_outputs: list[float], | ||
| sample_correction_factors: list[float], | ||
| recall_target: float, | ||
| precision_target: float, | ||
| delta: float | ||
| ) -> tuple[tuple[float, float], int]: | ||
| """Learns cascade thresholds given targets and proxy scores, | ||
| oracle outputs over the sample, and correction factors for the | ||
| sample.""" | ||
|
|
||
| def UB(mean, std_dev, s, delta): | ||
| return mean + (std_dev / (s ** 0.5)) * ((2 * np.log(1 / delta)) ** 0.5) | ||
|
|
||
| def LB(mean, std_dev, s, delta): | ||
| return mean - (std_dev / (s ** 0.5)) * ((2 * np.log(1 / delta)) ** 0.5) | ||
|
|
||
| def recall(pos_threshold: float, neg_threshold: float, sorted_pairs) -> bool: | ||
| helper_accepted = [x for x in sorted_pairs if x[0] >= pos_threshold or x[0] <= neg_threshold] | ||
| sent_to_oracle = [x for x in sorted_pairs if x[0] < pos_threshold and x[0] > neg_threshold] | ||
| total_correct = sum(pair[1] * pair[2] for pair in sorted_pairs) | ||
| recall = (sum(1 for x in helper_accepted if x[0] >= pos_threshold and x[1]) + sum(x[1] * x[2] for x in sent_to_oracle)) / total_correct | ||
| return recall | ||
|
|
||
| def precision(pos_threshold: float, neg_threshold: float, sorted_pairs) -> bool: | ||
| helper_accepted = [x for x in sorted_pairs if x[0] >= pos_threshold or x[0] <= neg_threshold] | ||
| sent_to_oracle = [x for x in sorted_pairs if pos_threshold > x[0] > neg_threshold] | ||
| oracle_positive = sum(x[1] for x in sent_to_oracle) | ||
| true_positives = sum(1 for x in helper_accepted if x[0] >= pos_threshold and x[1]) + oracle_positive | ||
| predicted_positives = sum(1 for x in helper_accepted if x[0] >= pos_threshold) + oracle_positive | ||
| precision = true_positives / predicted_positives if predicted_positives > 0 else 0 | ||
| return precision | ||
|
|
||
| # Pair helper model probabilities with helper correctness and oracle answer | ||
| paired_data = list(zip(proxy_scores, oracle_outputs, sample_correction_factors)) | ||
| sorted_pairs = sorted(paired_data, key=lambda x: x[0], reverse=True) | ||
| sample_size = len(sorted_pairs) | ||
|
|
||
| best_combination = (1,0) # initial tau_+, tau_- | ||
|
|
||
| # Find tau_negative based on recall | ||
| tau_neg_0 = max(x[0] for x in sorted_pairs[::-1] if recall(best_combination[0], x[0], sorted_pairs) >= recall_target) | ||
| best_combination = (best_combination[0], tau_neg_0) | ||
|
|
||
| # Do a statistical correction to get a new target recall | ||
| Z1 = [int(x[1]) * x[2] for x in sorted_pairs if x[0] >= best_combination[1]] | ||
| Z2 = [int(x[1]) * x[2] for x in sorted_pairs if x[0] < best_combination[1]] | ||
|
|
||
| mean_z1 = np.mean(Z1) if Z1 else 0 | ||
| std_z1 = np.std(Z1) if Z1 else 0 | ||
| mean_z2 = np.mean(Z2) if Z2 else 0 | ||
| std_z2 = np.std(Z2) if Z2 else 0 | ||
|
|
||
| corrected_recall_target = UB(mean_z1, std_z1, sample_size, delta/2)/(UB(mean_z1, std_z1, sample_size, delta/2) + LB(mean_z2, std_z2, sample_size, delta/2)) | ||
| corrected_recall_target = min(1, corrected_recall_target) | ||
| tau_neg_prime = max(x[0] for x in sorted_pairs[::-1] if recall(best_combination[0], x[0], sorted_pairs) >= corrected_recall_target) | ||
| best_combination = (best_combination[0], tau_neg_prime) | ||
|
|
||
| # Do a statistical correction to get a target satisfying precision | ||
| candidate_thresholds = [1] | ||
| for pair in sorted_pairs: | ||
| possible_threshold = pair[0] | ||
| Z = [int(x[1]) for x in sorted_pairs if x[0] >= possible_threshold] | ||
| mean_z = np.mean(Z) if Z else 0 | ||
| std_z = np.std(Z) if Z else 0 | ||
| p_l = LB(mean_z, std_z, len(Z), delta/len(sorted_pairs)) | ||
| if p_l > precision_target: | ||
| candidate_thresholds.append(possible_threshold) | ||
|
|
||
| best_combination = (max(best_combination[1], min(candidate_thresholds)), best_combination[1]) | ||
| oracle_calls = sum(1 for x in proxy_scores if best_combination[0] > x > best_combination[1]) | ||
|
|
||
| no_correction_sorted_pairs = [tup[:2] + (1,) for tup in sorted_pairs] | ||
| lotus.logger.info(f"Sample recall: {recall(best_combination[0], best_combination[1], no_correction_sorted_pairs)}") | ||
| lotus.logger.info(f"Sample precision: {precision(best_combination[0], best_combination[1], sorted_pairs)}") | ||
|
|
||
| return best_combination, oracle_calls | ||
|
|
||
| def calibrate_sem_sim_join(true_score: list[float]) -> list[float]: | ||
| true_score = np.clip(true_score, 0, 1) | ||
| return true_score |
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Just a simple question for my own understanding - is
assert stats["filters_resolved_by_helper_model"] > 0ever going to be flaky depending on the large/small model. Or is this super safe?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good question -- I think achieving >0 helper calls with a .9/.9 targets for this task should be simple enough for language models that I really wouldn't expect it to be flaky. Worst case if it comes to it the line can be removed, though I'd personally be a bit surprised if it failed.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for clarifying, let's keep it then