acnagle/Terminator-Qwen3-8B-Training-Data

Terminator-Qwen3-8B Training Data

Project page | Paper | Code | Trained weights

The tokenized, labeled data used to train Terminator-Qwen3-8B. Terminator is a small network that predicts when a reasoning model has reached its final answer so the remaining reasoning can be skipped.

Each row is one chain of thought (CoT) from Qwen3-8B on a math, coding or science problem, with per-token labels marking where the CoT first reaches its final answer. This dataset holds training data only. To evaluate on the test benchmarks, generate their CoTs with the code.

Contents

Source	Train CoTs	Train problems	Validation CoTs	Validation problems
MATH	27,666	10,745	134	50
OpenCoder	22,875	9,423	109	50
OpenScience	29,933	11,719	123	50
AIME 1983-2024	1,625	751	109	50
Total	82,099	32,638	475	200

Usage

from datasets import load_dataset
from transformers import AutoTokenizer

ds = load_dataset("acnagle/Terminator-Qwen3-8B-Training-Data")
row = ds["train"][0]

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
text = tokenizer.decode(row["input_ids"])  # the prompt and the CoT

To train with the code, download the files to the path its scripts expect, then run the training script from the repo root:

hf download acnagle/Terminator-Qwen3-8B-Training-Data train.jsonl val.jsonl \
  --repo-type dataset --local-dir data/outputs/Qwen/Qwen3-8B
bash train_terminator_qwen3-8b.sh

Fields

Field	Description
cot_idx	CoT ID, unique across both splits
sample_idx	Problem ID, specific to this dataset. CoTs of the same problem share it
problem, answer	Problem text and reference answer
solution	Reference solution. Null for MATH (see Sources and licenses) and AIME (none upstream)
data_source	math, opencoder, openscience or aime_<year>
data_fields	Source-specific metadata. Keys that don't apply to a source are null
input_ids	Qwen3-8B tokens: prompt, <think>, the CoT, </think>. The final response isn't included
labels	-100 outside the reasoning, 0 up to and including the last token of the answer span, 1 after it. </think> is always 1
valid_mask	1 where labels isn't -100. The training code doesn't read it

How it was made

Each problem has 3 CoTs, sampled from Qwen3-8B with temperature 0.6, top-p 0.95, top-k 20 and presence penalty 1.5, up to 24,576 tokens. Qwen3-30B-A3B-Thinking-2507 then extracted each final answer and located the first place the CoT reaches it. Only CoTs whose span passed verification were kept: 82,574 of 110,345, covering 32,838 of 36,933 problems.

prepare_training_data.py tokenized those CoTs and built the labels with its default settings, holding out 50 problems per source for validation using seed 1337. The rows are grouped by source, so shuffle them before training.

Rebuilding this data from scratch requires generating CoTs and rerunning the pipeline in the code.

Sources and licenses

CoTs, spans and labels are released under Apache-2.0, matching Qwen3-8B. Problem text and reference data stay under their original terms. The labels below are as listed upstream, and they may not settle the rights to the underlying problems.

Source	License listed upstream
MATH, via the PRM800K split	None. solution is omitted because AoPS claims those solutions as proprietary
AIME 1983-2024	MIT
OpenCoder (educational_instruct)	MIT
OpenScience (OS-Q3-235B-4)	CC BY 4.0

Citation

@misc{nagle2026terminatorlearningoptimalexit,
  title   = {TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning},
  author  = {Alliot Nagle and Jakhongir Saydaliev and Dhia Garbaya and Michael Gastpar and Ashok Vardhan Makkuva and Hyeji Kim},
  year    = {2026},
  eprint  = {2603.12529},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url     = {https://arxiv.org/abs/2603.12529}
}