Add mined-vs-random negatives experiment with measured results#10
Merged
Conversation
banking77, Qwen2.5-0.5B + LoRA, identical per-arm budgets, RTX 4080: self-mined pools beat random negatives by +5.0 MAP@25 / +8.6 R@1, with the gain concentrated at top-1 where sibling labels collide. The --cold-start ablation shows mining from an untrained model collapses to MAP 0.430 (below random) - hard negatives are only as good as their miner, which is why the pipeline bootstraps on random negatives first (the competition protocol).
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Measures the library's central claim end to end through its public API on a public dataset (banking77, 77-intent closed bank; Qwen2.5-0.5B-Instruct + LoRA; 2,000 train / 1,000 held-out; pools of 8; identical budgets per arm; one RTX 4080, ~1 h):
--cold-startablation: mining round 1 from the zero-shot model instead of the bootstrap model collapses to MAP 0.430 — far below random. Hard negatives are only as good as their miner; the experiment mirrors the competition protocol (bootstrap → mine from the previous round's model, fresh adapter per round), and the README now states this caveat with numbers.🤖 Generated with Claude Code