Echo Prompt Minimal Share

This repository is a small, shareable snapshot for the paper "Mirror, Mirror in the Wall..." submitted to EMNLP.

Contents

Prompt datasets live in prompt_datasets/.

Dataset	File	Description
A	dataset_A_self_generated_echo_384.jsonl	Original self-generated echo-screened prompts, 64 prompts from each of six source models.
B	dataset_B_iterated_self_prompt_echo_384.jsonl	Echogenic dataset 2. Each prompt is the final answer from a 10-step answer-as-next-prompt chain, again 64 per source model.
C	dataset_C_non_echo_reference_128.jsonl	Non-echo reference prompts. These passed a cross-model screen with zero prefix partial/full echo.

Model metadata is in configs/models.json. The scripts assume they are run from the repository root. On Slurm, override VENV_DIR, HF_HOME, and MODEL_KEY_ARGS as needed for the target cluster.

Experiment 1: Echo Evaluation

Experiment 1 evaluates non-greedy continuations of each prompt and counts prefix full echo and prefix partial echo. Partial echo means that the answer starts as an exact prompt copy for at least 40 percent of the prompt tokens, but does not copy the whole prompt.

PROMPTS_JSONL=prompt_datasets/dataset_A_self_generated_echo_384.jsonl \
RUN_ID=exp1_dataset_A \
sbatch scripts/run_experiment1_echo_eval.sbatch

PROMPTS_JSONL=prompt_datasets/dataset_B_iterated_self_prompt_echo_384.jsonl \
RUN_ID=exp1_dataset_B \
sbatch scripts/run_experiment1_echo_eval.sbatch

Direct Python entrypoint:

python scripts/run_echo_eval.py \
  --prompts-jsonl prompt_datasets/dataset_A_self_generated_echo_384.jsonl \
  --models-config configs/models.json \
  --output-jsonl outputs/experiment1/example/completions.jsonl \
  --summary-json outputs/experiment1/example/summary.json \
  --summary-md outputs/experiment1/example/summary.md

Experiment 2: Training Data Overlap

The latest Experiment 2 only scans the relevant echo-confirmed prompts: OLMo-echoing prompts are compared with the OLMo/Tulu-3 SFT mixture, and SmolLM2-echoing prompts are compared with SmolTalk. First build those filtered prompt sets from Experiment 1 completions:

python scripts/build_experiment2_filtered_prompts.py \
  --completion-jsonl outputs/experiment1/exp1_dataset_A/completions.jsonl \
  --completion-jsonl outputs/experiment1/exp1_dataset_B/completions.jsonl \
  --output-dir outputs/experiment2_filtered

Then scan the filtered echo prompts:

PROMPTS_JSONL=outputs/experiment2_filtered/olmo_echo_prompts_datasetA_B.jsonl \
RUN_ID=exp2_olmo_echo_tulu3 \
DATASET_PRESETS=',' \
DATASET_SPECS='olmo_tulu3|allenai/tulu-3-sft-olmo-2-mixture|default|train|' \
SCAN_MODE=initial_words \
INITIAL_WORD_THRESHOLDS=3,5 \
sbatch scripts/run_training_overlap_scan.sbatch

PROMPTS_JSONL=outputs/experiment2_filtered/smol_echo_prompts_datasetA_B.jsonl \
RUN_ID=exp2_smol_echo_smoltalk \
DATASET_PRESETS=',' \
DATASET_SPECS='smoltalk|HuggingFaceTB/smoltalk|all|train|' \
SCAN_MODE=initial_words \
INITIAL_WORD_THRESHOLDS=3,5 \
sbatch scripts/run_training_overlap_scan.sbatch

For the non-echo control, run the same scans with Dataset C:

PROMPTS_JSONL=prompt_datasets/dataset_C_non_echo_reference_128.jsonl \
RUN_ID=exp2_nonecho_tulu3 \
DATASET_PRESETS=',' \
DATASET_SPECS='olmo_tulu3|allenai/tulu-3-sft-olmo-2-mixture|default|train|' \
SCAN_MODE=initial_words \
INITIAL_WORD_THRESHOLDS=3,5 \
sbatch scripts/run_training_overlap_scan.sbatch

PROMPTS_JSONL=prompt_datasets/dataset_C_non_echo_reference_128.jsonl \
RUN_ID=exp2_nonecho_smoltalk \
DATASET_PRESETS=',' \
DATASET_SPECS='smoltalk|HuggingFaceTB/smoltalk|all|train|' \
SCAN_MODE=initial_words \
INITIAL_WORD_THRESHOLDS=3,5 \
sbatch scripts/run_training_overlap_scan.sbatch

Optional role localization for matched instruction datasets:

python scripts/analyze_overlap_match_roles.py \
  --run smol_echo=outputs/training_overlap/exp2_smol_echo_smoltalk/best_matches.jsonl \
  --run olmo_echo=outputs/training_overlap/exp2_olmo_echo_tulu3/best_matches.jsonl \
  --output-json outputs/training_overlap/roles_summary.json \
  --output-md outputs/training_overlap/roles_summary.md

Experiment 3: Forced Seed Echo

Experiment 3 starts from non-echo reference prompts and simulates a model that has already begun responding with the first k prompt tokens. Run separate jobs for k = 2, 3, and 5.

for k in 2 3 5; do
  PROMPTS_JSONL=prompt_datasets/dataset_C_non_echo_reference_128.jsonl \
  SEED_TOKEN_COUNT=$k \
  RUN_ID=exp3_seed${k} \
  sbatch scripts/run_experiment3_seeded_echo_eval.sbatch
done

Experiment 4: Copy-Head Ablation Echo Evaluation

Experiment 4 repeats the Experiment 1 echo evaluation for HF/Transformers models while ablating heads with high repeated-random-sequence copying score. The default conditions are baseline, head_ablation_top8, head_ablation_top16, and head_ablation_random16. Direct head ablation is only implemented for HF/Transformers models; GGUF and LiteRT quantized backends do not expose the same attention-head hooks in these scripts.

PROMPTS_JSONL=prompt_datasets/dataset_A_self_generated_echo_384.jsonl \
RUN_ID=exp4_dataset_A \
sbatch scripts/run_experiment4_head_ablation_echo_eval.sbatch

PROMPTS_JSONL=prompt_datasets/dataset_B_iterated_self_prompt_echo_384.jsonl \
RUN_ID=exp4_dataset_B \
sbatch scripts/run_experiment4_head_ablation_echo_eval.sbatch

The copy-head ranking is computed inside each run using INDUCTION_PROBE_TRIALS=8 repeated random-token probes of length INDUCTION_PROBE_SEQ_LEN=64 by default.