Conversation
…th it
One forward pass over [prompt][MASK x c], before any token exists, predicts how long the
answer will be. Compared against DAEDAL's Stage-1, which finds the same number by growing
the canvas ~40 forward passes at a time, and against SmartCrop, which like the probe costs
one pass.
Only probe_outlen.py is new code. The rest already existed:
capture capture_task_features.py rebuilds THIS repo's prompt (task fn -> chat template
-> assistant prefill) and reads the dataset from the task YAML, so doc_id here
is the harness doc_id. Features captured under another prompt cannot be joined
to labels measured under this one.
label metrics/math500.py already writes per_problem_tokens.csv; the informative column
is effective_tokens, not total_raw_tokens, which is just the canvas.
run scripts/eval_LLaDA_ProbeLen.sh
Every arm runs the SAME decoder and differs only in the initial canvas. The baseline script
header records why: the vanilla LLaDA decoder and the DAEDAL decoder produced 0/34 identical
outputs at canvas 32 on the same problems, so a policy-vs-uniform delta measured across
decoders would mix allocation with decoding.
alloc_out on LLaDA_DAEDAL is additive and exists so Stage-1's prediction can be read per
problem -- scored as a predictor, not only as an end-to-end policy. SmartCrop already wrote
the same schema, so probe_outlen.py --compare reads all three from one function.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Five steps, in dependency order, with the two checks that decide whether to continue: the truncation rate at canvas 2048 (a truncated answer is a lower bound, not a length -- at 1024 in the companion repo 55% of MATH answers were still running), and the Experiment 1 table, which gates whether Experiment 2 is worth the GPU time. Also makes TASKS overridable in eval_LLaDA_Baseline.sh. It was hardcoded to math500, and the probe is fitted on math_train, so labelling the fit set needed either this one-line change or a second near-identical script. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Not every dataset has a train split, and gsm8k/humaneval/mbpp have none here. --cv K fits out-of-fold on the eval task itself using make_cv_folds copied from dllm_probe/probe/cv.py -- stratified by level, 20%% test, 60/20 train/val, seed 0 -- so answer length is estimated under the same protocol as difficulty and success in the companion repo. CV is free in GPU terms: the expensive step is generation, not fitting, and out-of-fold predictions concatenate into one allocation CSV that a single generation run consumes. Every row was predicted by a model that never saw that problem. fit() now takes explicit train/val indices instead of splitting internally, which is what lets one function serve both the fixed-split and per-fold paths. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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.
No description provided.