README.md

metadata

pretty_name: DR-AntiForget RQ3 Writing Refresh
tags:
  - continual-learning
  - lora
  - evaluation

RQ3 Writing Refresh: Reproducibility Artifacts

This directory contains the compact reproducibility package for the RQ3.1 experiment that replaced only the lora-writing adapter with an OpenScholar-refreshed adapter (writing++) while keeping the Qwen3-32B base, Plan adapter, and Search adapter frozen.

Headline result

The refresh improved teacher-forced performance on the new OpenScholar domain but did not yield a reliable general downstream improvement:

Evaluation set	Source PPL	Refreshed PPL	Relative change
OpenScholar holdout	3.5567	2.7634	-22.3%
Historical Writing holdout	1.5866	1.6561	+4.4%

The downstream table should be interpreted as a preliminary negative/mixed result. Every paired 95% bootstrap confidence interval includes zero. See analysis/DECLINE_ANALYSIS.md for the failure and routing audit.

Contents

• model/lora-writing/: refreshed adapter config and weights. The weight SHA-256 is 4e31f83c31f4784647147eea5458f5e96cd7712f5a88b21c6fd1ae2b9919e6c5.
• downstream/downstream_examples.jsonl: 1,800 final answers with questions, scores, failure flags, and source-file provenance; 2 adapter conditions × 3 routing modes × 3 benchmarks × 100 tasks.
• downstream/aggregate_results.json, table.md, and table.tex: exact aggregate reproduced from the 1,800 exported rows.
• analysis/analysis_summary.json: paired bootstrap intervals, win/tie/loss counts, answer lengths, failure transitions, dataset-shift statistics, and stage-reach summaries.
• analysis/paired_deltas.csv: all 21 routing/benchmark/metric paired comparisons.
• analysis/trajectory_diagnostics.jsonl: compact per-task routing diagnostics without raw chain-of-thought or judge reasoning.
• analysis/dataset_shift.json: descriptive statistics for the new-domain train split and historical Writing holdout.
• splits/openscholar_split_manifest.json: exact trajectory-grouped split (seed 42), including 795 train rows and 205 holdout rows with zero group overlap.
• metadata/: training state, paired NLL/PPL evaluation, experiment metadata, model/data revisions, and adapter hashes.
• scripts/build_export.py: deterministic export and validation script. It selects the latest scored record per cell/task, verifies all 18 cells against the published aggregate, and builds the compact files above.

Existing source data

The raw source datasets already live in WeAct/DR-AntiForget and are not duplicated here:

• lora_writing_openscholar/lora-writing-openscholar.jsonl (1,000 rows)
• tinker_holdout_test_v2/lora-writing_holdout_test.jsonl (200 historical holdout rows)

Dataset revision used by training: 3dde83eb8ee2518c3eea4ef5317a40189215e475.

JSONL schema

Each row in downstream/downstream_examples.jsonl has:

• condition: original_writing or writing_pp
• routing: hard, soft, or soft_oracle
• benchmark: researchqa, healthbench, or deepresearchbench
• task_id, task_question, answer, answer_source, answer_length
• framework_failure: answer begins with Task incomplete or Task interrupted
• metrics: raw 0–1 judge metrics (coverage, score, or the four DRB dimensions plus their mean)
• source_files: paths relative to the source evaluation run

Caveats

HealthBench and DeepResearchBench scores use preliminary local Qwen3-32B judges. The run has one training seed and independently sampled agent/search trajectories. Use the retained answers for official-judge rescoring, and do not treat the current downstream deltas as statistically significant or as a clean causal estimate of the Writing adapter effect.