nur-dev/stress17k-counterfactual-pairs

Beyond Transcript Alignment — derived data

Derived data + per-seed evaluation results for the Beyond Transcript Alignment research project (frozen-frozen speech-to-LLM adapters under counterfactual training on the StressTest benchmark).

Code: https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress

Contents

• cf_pairs/cf_pairs_train.jsonl — 3666 same-transcript counterfactual pairs from Stress-17K-raw probe-train (transcript IDs + stress indices; no raw audio).
• cf_pairs/cf_pairs_artifact.jsonl — 6846 artifact-matched negatives (same Φ, different surface features).
• cf_pairs/cf_pairs_train_shuffled.jsonl — Stage 4 Control B: 3666 pairs with audio decorrelated from (transcript, Φ) labels.
• proj_P/proj_P.pt — Stage 4 Control A: 4096→1024 lstsq-fit linear projection (train MSE 0.090, train cos-sim 0.999) used to construct the text-only adapter baseline.
• desc_only_baseline.json — pinned cross-domain absolute Probe-K floor (0.361).
• eval_results/stage{2,3p8,4,6,6_r4,7}_eval/ — per-seed summary.json
  • rows.json files for every cohort and pilot in the project.

Audio access

No raw audio is included. Audio for the four datasets used in this project is accessed via HuggingFace Hub:

• slprl/StressPresso (CC-BY-NC-4.0) — n=202 test items
• slprl/Stress-17K-raw (CC-BY-NC-4.0) — cf-pairs derive from this
• ylacombe/expresso (CC-BY-NC-4.0) — artifact-matched negatives
• openslr/librispeech_asr config clean (CC-BY-4.0) — domain mix

The cf-pairs JSONL files contain transcription_id + stress_index + speaker_id + style references; users instantiate audio at load time from the upstream HuggingFace datasets.

Reproducibility

from huggingface_hub import snapshot_download
import json

# Download
local = snapshot_download("nur-dev/stress17k-counterfactual-pairs", repo_type="dataset")

# Reproduce R1.8 cohort Probe-G_neutral (= 0.5122 ± 0.0039 across 5 seeds)
import numpy as np
seeds = [1234, 1235, 1236, 1237, 1238]
vals = [json.load(open(f"{local}/eval_results/stage3p8_eval/seed{s}/summary.json"))
            ["adapter"]["accuracy_neutral"] for s in seeds]
print(f"R1.8 Probe-G_neutral cohort: {np.mean(vals):.4f} ± {np.std(vals):.4f}")

License

CC-BY-NC-4.0 (inherited from Stress-17K-raw upstream). Permitted for academic research, ablation studies, reproducibility checks, pedagogy. Commercial use is not permitted under upstream dataset licenses.

Citation

@software{kadyrbek_frozen_stress_llm,
  author  = {Kadyrbek, Nurgali},
  title   = {frozen-speech-llm-stress: research code for frozen-frozen speech-to-{LLM} adapters under counterfactual training},
  year    = {2026},
  url     = {https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress}
}