Beaverdam/cliff-token-data

Cliff Tokens: Stem Traces and Token-Level Rollouts

Data for Cliff Tokens: Analyzing Failure Trigger Tokens in LLM Mathematical Reasoning. Source code and analysis instructions: Cliff-token and download script.

Project page: https://jaeyongko.github.io/cliff-token/

The files are uncompressed JSON lists. Each model and dataset has a separate file in both directories:

• stem_traces/<model>/<dataset>.json: one sampled response per problem, including token IDs and correctness.
• rollouts/<model>/<dataset>.json: the same response with all_position_scores, the estimated probability of a correct answer after each token based on 64 sampled continuations.

all_position_scores[i] corresponds to the prefix response_token_ids[:i+1]; the list has one fewer element than response_token_ids.

Models: Qwen3-0.6B, Qwen3-4B, Qwen3-8B, Llama-3.2-1B-Instruct, Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, gemma-3-4b-it.

Datasets: gsm1k (100), math500 (100), and aime25 (30) for every model; gsm8k_train (7,473) for Qwen3-0.6B, Llama-3.2-1B-Instruct, and Llama-3.1-8B-Instruct.

import json
from pathlib import Path

traces = json.loads(Path("rollouts/Qwen3-8B/math500.json").read_text())
print(traces[0]["all_position_scores"])

Fields

Field	Meaning
id, problem_id	Trace and problem identifiers
question, golden_answer	Problem and accepted answers
full_prompt	Model input after the chat template
response	Generated stem trace
response_tokens, response_token_ids	Response tokens and token IDs
total_tokens	Number of response tokens
is_correct	Whether the final answer is correct
all_position_scores	Rollout success probabilities; only in rollouts/

Data is released under CC BY 4.0. Benchmark problems and model outputs remain subject to the original dataset licenses and model terms of use.