metacognitive-behavioral-tuning/mbt-traces

🧠 MBT Traces: teacher reasoning with inserted metacognitive behaviors

This dataset accompanies the paper "Distilling Metacognitive Behaviors through Agentic Editing" (anonymous submission). It contains 1,000 competition-style math problems from Nemotron-Math-v2, the original reasoning traces of the teacher model DeepSeek-V4.1-Flash on those problems, and the same traces after three editing strategies inserted seven metacognitive behaviors into them:

Strategy	What the editor does
Unedited	Nothing. The teacher's trace as generated.
MBT-R (rewrite)	Rewrites the whole trace in one call, returning the full edited reasoning and response.
MBT-I (insert)	Sees the reasoning with a numbered marker after each paragraph and returns, in one call, a list of passages to insert after chosen markers.
MBT-A (agentic)	A coding agent edits the trace as a file in a git repository, inserting one passage per commit.

The seven target behaviors are subgoal setting, logical check, substantive verification, targeted backtracking, resuming abandoned paths, self-correction and sufficiency judgment. The companion GitHub repository contains the generation and editing code, the exact prompts, and both annotation schemes.

Models fine-tuned from Qwen3.5-4B-Base on each condition: Qwen3.5-4B-Unedited-SFT, Qwen3.5-4B-MBT-R, Qwen3.5-4B-MBT-I and Qwen3.5-4B-MBT-A.

📦 Contents

Config	Rows	What a row is
problems	1,000	A problem, its reference answer, its difficulty bin and its source attribution
traces_unedited	1,000	The teacher's reasoning and response for one problem
traces_mbt_r	1,000	The MBT-R edit of that trace (776 edited, 224 unsuccessful)
traces_mbt_i	1,000	The MBT-I edit (1,000 edited, 6,472 insertions)
traces_mbt_a	1,000	The MBT-A edit (1,000 edited, 10,759 insertions)
sft_unedited	995	Chat-format training examples for Unedited SFT
sft_mbt_r	773	Chat-format training examples for MBT-R
sft_mbt_i	995	Chat-format training examples for MBT-I
sft_mbt_a	995	Chat-format training examples for MBT-A

sft_* configs

Each condition is trained separately on its own config. Each keeps the traces that were edited successfully and fit within the student's maximum length of 131,072 tokens. Token counts are under the Qwen3.5-4B chat template (paper Table D1).

Config	Examples	Tokens (Table D1)
sft_unedited	995	22,821,876
sft_mbt_r	773	7,557,835
sft_mbt_i	995	23,483,410
sft_mbt_a	995	24,432,120

🗂️ Fields

problems

Field	Type	Description
id	string	Problem id, shared by every file
problem	string	The problem statement as given to the teacher
reference_answer	string	The reference answer from Nemotron-Math-v2 (never shown to any editor)
difficulty_bin	int	0 to 8: how many of eight gpt-oss-120b attempts were correct
source	string	stackexchange or aops
license	string	CC-BY-SA-4.0 (StackExchange) or CC-BY-4.0 (AoPS)
source_url, author_name, author_url	string or null	Attribution as given by Nemotron-Math-v2 (StackExchange rows)

traces_unedited: id, problem, reasoning (the text inside the thinking block), response (the text after it), finish_reason.

traces_mbt_r, traces_mbt_i, traces_mbt_a: id, method, edited (false when the strategy produced no acceptable edit), reasoning, response (null when not edited), response_changed (the editor changed the response section), and insertions, a list of {behavior, position, text} (MBT-I also gives marker). position is the share of the original reasoning, counted in non-whitespace characters, that comes before the insertion: 0 is the start and 1 the end. MBT-I positions follow from the markers; MBT-A positions come from replaying the agent's commits. MBT-R rewrites the whole trace, so it has no insertion list.

sft_unedited, sft_mbt_r, sft_mbt_i, sft_mbt_a: id and messages, a user message (the problem followed by "Please reason step by step, and put your final answer within \boxed{}.") and an assistant message <think>\n{reasoning}\n</think>\n\n{response}. Each config is one condition's training data (Table D1).

🚀 Usage

from datasets import load_dataset

problems = load_dataset("metacognitive-behavioral-tuning/mbt-traces", "problems", split="train")
mbt_a = load_dataset("metacognitive-behavioral-tuning/mbt-traces", "traces_mbt_a", split="train")
sft_mbt_a = load_dataset("metacognitive-behavioral-tuning/mbt-traces", "sft_mbt_a", split="train")

🛠️ How the data was made

1. Problems. A pool of 10,000 Nemotron-Math-v2 problems after deduplication and decontamination against the five evaluation benchmarks (AIME 2025, AIME 2026, HMMT February 2025, HMMT February 2026 and HMMT November 2025) and seven more math benchmarks (MATH-500, AMC 2023, AIME 2024, HMMT February 2023, HMMT February 2024, OlympiadBench and GSM8K). The pool is divided into nine difficulty bins by how many of eight gpt-oss-120b attempts are correct, and 111 or 112 problems are drawn per bin.
2. Teacher. DeepSeek-V4.1-Flash solves each problem once (reasoning effort max, temperature 1.0, top_p 0.95). Incorrect traces are kept.
3. Editing. The editor is the teacher model itself. It sees the problem, the reasoning and the response, never the reference answer, and may not look the problem up. All three strategies share the same behavior instructions.

⚖️ License and attribution

The dataset is released under CC BY-SA 4.0 (see LICENSE.md), following the license of the StackExchange subsets of Nemotron-Math-v2. It contains only Nemotron-Math-v2 problems and outputs of DeepSeek-V4.1-Flash. See NOTICE.md for the upstream licenses and the per-problem attribution.

⚠️ Limitations

• The traces are from one teacher on one domain (math). Incorrect teacher answers are retained; editing does not guarantee a correct answer.
• Inserted passages are model-written. They are meant to demonstrate habits of checking and revising, and individual passages can themselves contain errors.
• StackExchange problems are user-contributed and may contain ambiguities.

📚 Citation

@misc{mbt2026,
  title  = {Distilling Metacognitive Behaviors through Agentic Editing},
  author = {Anonymous},
  year   = {2026},
  note   = {Under review}
}