TopoAgent Planner SFT Data (910 examples)
This folder contains the 910-row parquet dataset used for the TopoAgent planner SFT runs. It was generated from the BrowseComp-Plus training split by filtering GPT-4.1 teacher-generated DAG planning trajectories. The validation split was separate and was not used to generate these samples.
The file has 910 records covering 304 distinct question_idx values. Each record contains:
messages: a multi-turn list of role/content messages for supervised fine-tuning;question_idx: the source training question index;round: the trajectory round represented in the example;reward: the retained-example reward label (1 for all published rows).
The parquet is a compact training artifact; it does not include the BrowseComp-Plus corpus, search index, teacher API logs, or model checkpoints. It contains training questions and generated trajectories, so use it only as training data. Do not use these examples as an evaluation set or report scores on them as held-out results.
Load
from huggingface_hub import hf_hub_download
from datasets import load_dataset
path = hf_hub_download(
repo_id="anonymous2222222/tpa",
filename="TopoAgent-SFT-Data-910/sft_samples.parquet",
)
dataset = load_dataset("parquet", data_files=path, split="train")
print(len(dataset)) # 910
For the SFT recipe and model weights, see TopoAgent-4B-SFT and TopoAgent-8B-SFT. The training recipe used 910 examples, LoRA rank 32 / alpha 64, 3 epochs and 84 optimizer steps. The training and validation files were the same in that historical SFT run; its reported SFT validation loss is not a held-out generalization estimate.
Source and benchmark note
The source task split is Tevatron/BrowseComp-Plus. The upstream BrowseComp project requests that examples not be posted online in plaintext to reduce benchmark leakage. This artifact contains training-split examples and is placed in a separate directory in the existing model repository at the user's request. See the upstream BrowseComp notice and BrowseComp-Plus dataset card.
The original teacher rollout records and exact generation/filtering run configuration are not bundled. This is the resulting filtered SFT file, not a claim of full end-to-end data-generation reproducibility.