Skip to content

README.md

Latest commit

 

History

History
129 lines (80 loc) · 2.88 KB

README.md

File metadata and controls

129 lines (80 loc) · 2.88 KB
Raw

Classifier Experiment

This folder contains the full pipeline for the first experiment in the paper: Probing with Classifier

The workflow is controlled by one config file:

  • configs/classifier.yaml

Each entry script reads only its own config section (classifier_data, combine_label, classifier_train, correctness_eval, statistic_baseline).

Pipeline Overview

  1. Generate classifier data (.npz) and LLM judgment prompts (*_judge.jsonl).
  2. Call external LLM API (not included in this repo) on *_judge.jsonl.
  3. Merge LLM labels (Correctness, Logicality) back into .npz.
  4. Train classifiers for different targets.
  5. Evaluate Correctness/Logicality classifiers.
  6. Compute statistic baselines.

Step 0: Configure

Edit config first:

vim configs/classifier.yaml

Step 1: Generate classifier data

Script:

bash scripts/run_classifier_data.sh configs/classifier.yaml

What it produces:

  • *.npz with fields such as latents, hints, step_lengths, begin_token_ids, correctness, logicality
  • *_judge.jsonl prompts for LLM-based correctness/logicality labeling

Optional override example:

bash scripts/run_classifier_data.sh configs/classifier.yaml \
  --set checkpoint_name=gsm8k-385k_Qwen2.5-0.5b_spar-10.pt \
  --set input_file=data/gsm8k_385K_valid.json \
  --set output_file=classifier_data/gsm8k_385K_valid_classifier_data

Step 2: Call LLM API externally

Input:

  • the *_judge.jsonl file from Step 1

Expected API output JSONL (at least):

  • id
  • response (or a custom response field)

Step 3: Merge Correctness/Logicality labels

Command:

PYTHONPATH=. python classifier/combine_label.py --config configs/classifier.yaml

If your response text field is not response, set the response_key manually:

PYTHONPATH=. python classifier/combine_label.py --config configs/classifier.yaml --set response_key=content

Step 4: Train classifiers

Script:

PROC=<num_gpus> bash scripts/run_classifier_train.sh configs/classifier.yaml

You can set different inputs and targets by changing classifier_train.task:

inputs:

  • Tr
  • hints

targets:

  • len
  • token
  • correctness
  • logicality

Notes:

  • For step length and first token ppl, training loss is used directly as the main evaluation signal.
  • For correctness and logicality, run Step 5 evaluation.

Step 5: Evaluate Correctness/Logicality

Script:

bash scripts/run_correctness_eval.sh configs/classifier.yaml

This reports accuracy for both correctness and logicality classifiers.

Step 6: Statistic baselines

Command:

PYTHONPATH=. python classifier/statistic_baseline.py --config configs/classifier.yaml

This reports baselines for all four labels:

  • step length (mean/std)
  • first token ppl (entropy/perplexity)
  • correctness (positive ratio)
  • logicality (positive ratio, when enabled)