AZH04/wrong-reasoning-traces

Wrong-Reasoning Trace Bank (Completeness Cliff)

Model-generated reasoning traces used as the injection material for the pilot study "The completeness cliff: language models escape wrong reasoning until it's finished." Each row is one sampled baseline rollout, kept so its reasoning can be spliced back into a fresh prompt to test whether a model can still recover the correct answer (pass@k).

• Code and write-up: https://github.com/AlvinZH04/Completeness-Cliff
• Blog post: https://alvinzh04.github.io/blog/completeness-cliff.html

What is here

One JSONL file per baseline run (data/base_<model>_<dataset>.jsonl). The correct == false rows are the self_wrong injection source (the model's own wrong reasoning); the correct == true rows feed the corrupted source (a correct scaffold with numbers perturbed near the cut) and serve as irrelevant / cross-domain donors for other questions.

Schema

field	type	meaning
run	string	baseline run this rollout came from
qid	string	question id (e.g. aime25-3)
dataset	string	aime24_25, rg_maze, rg_mini_sudoku
rg_task	string/null	reasoning-gym task name, if applicable
gold	string	correct answer
sample_index	int	which of the N=16 samples
answer	string	the model's extracted answer for this sample
correct	bool	whether answer matched gold
truncated	bool	hit the generation length cap
n_tokens	int	generated token count
trace_text	string	the reasoning channel (a reasoning model's <think> content; an instruct model's response). This is what gets injected.

Models and problems

model	kind	AIME 24+25	reasoning-gym
Qwen3-4B-Thinking-2507	reasoning	yes	maze, mini-sudoku
Qwen3-4B-Instruct-2507	instruct (matched sibling)	yes (+ a 32k-budget rerun)	maze
Gemma-4-E2B-it	hybrid (cross-family)	yes	maze

N = 16 samples per question, official card-recommended sampling, vLLM. Per-run baseline metrics (pass@k, adoption, truncation, token counts) are in baseline_summaries.json.

Load

from datasets import load_dataset
ds = load_dataset("AZH04/wrong-reasoning-traces", split="train")
wrong = ds.filter(lambda r: not r["correct"])   # the self_wrong injection traces

Or a single run directly:

import json
rows = [json.loads(l) for l in open("data/base_qwen3-4b-thinking_aime24_25.jsonl")]

Notes

• Only distilled fields are kept. The full raw rollouts (with the post-think answer text and token-level metadata, ~2.4 GB) are not published here; this bank preserves the injectable reasoning and is regenerable from the code repo.
• Question text is not redistributed; rows reference AIME problems by qid and include the gold answer only. AIME problems are the property of the Mathematical Association of America.
• Traces are outputs of Qwen3 (Apache-2.0) and Gemma (Gemma Terms of Use) models. The dataset card and structure are released CC-BY-4.0; model outputs are subject to the source models' terms.

Citation

If you use this, please cite the project:

@misc{zhang2026completenesscliff,
  title        = {The completeness cliff: language models escape wrong reasoning until it's finished},
  author       = {Zhang, Alvin},
  year         = {2026},
  howpublished = {\url{https://alvinzh04.github.io/blog/completeness-cliff.html}},
  note         = {wrong\_reason notes}
}