bactrianus/bactrainus-hotpotqa-teacher-traces

Bactrainus HotpotQA Teacher Traces

SOURCE-LINKED v1.0

Archived Llama 3.1 rationale and question-decomposition supervision, paired with complete SFT conversations and stable HotpotQA identities.

198,660 ROWS 4 CONFIGURATIONS SFT MESSAGES 8B + 70B LABELS CC BY-SA 4.0

A focused release of recovered teacher-generated supervision for multi-hop question answering. Every row contains the normalized annotation, an ordered system/user/assistant conversation, the gold source fields needed to interpret it, and a stable source_id that joins to the canonical Bactrainus HotpotQA release.

Trace + conversationPlain annotation text and training-ready messages are published together.

Stable source joinOfficial HotpotQA IDs, verified source rows, gold context, evidence, and artifact digests.

Resolved recordsOne output per source ID; a complete rerun replaces its older base output on overlap.

At a glance

Property	Value
Dataset	bactrianus/bactrainus-hotpotqa-teacher-traces
Release	v1.0.0
Tasks	Evidence-grounded rationale generation and question decomposition
Configurations	4
Rows	198,660 across configurations
Parquet shards	42
Split	train only
Canonical join	bactrianus/bactrainus-hotpotqa at v1.0.0
Upstream source	hotpotqa/hotpot_qa, distractor/train
Join key	source_id
Reported teacher labels	Llama 3.1 8B Instruct and Llama 3.1 70B Instruct
Storage	Zstandard-compressed Parquet
License	CC BY-SA 4.0

This repository contains training annotations only. It does not include development or test examples, model predictions, benchmark scores, evaluation files, or result tables.

Quick start

Install 🤗 Datasets and choose the task and reported teacher size explicitly:

from datasets import load_dataset

traces = load_dataset(
    "bactrianus/bactrainus-hotpotqa-teacher-traces",
    "teacher-cot-llama31-70b-sft",
    split="train",
    revision="v1.0.0",
)

row = traces[0]
print(row["source_id"])
print(row["annotation"])
print(row["messages"])

The rows are already serialized as SFT conversations. Apply the chat template of the model being trained; the dataset intentionally stores no model-specific control tokens:

rendered = tokenizer.apply_chat_template(
    row["messages"],
    tokenize=False,
    add_generation_prompt=False,
)

Join a trace to the complete canonical candidate context by source_id:

source = load_dataset(
    "bactrianus/bactrainus-hotpotqa",
    "structured",
    split="train",
    revision="v1.0.0",
)

source_by_id = {example["source_id"]: example for example in source}
canonical_example = source_by_id[row["source_id"]]

For exploratory access without downloading a complete configuration:

stream = load_dataset(
    "bactrianus/bactrainus-hotpotqa-teacher-traces",
    "teacher-decomposition-llama31-8b-sft",
    split="train",
    revision="v1.0.0",
    streaming=True,
)
first = next(iter(stream))

Configuration guide

Each configuration contains one non-empty resolved annotation per source_id. Counts across configurations are not counts of globally distinct HotpotQA questions: the same source question may legitimately appear under different tasks or teacher labels.

Configuration	Task	Reported teacher label	Easy	Medium	Hard	Rows	Source coverage
teacher-cot-llama31-8b-sft	Rationale + answer	Llama 3.1 8B Instruct	14,037	44,971	12,230	71,238	78.7621%
teacher-cot-llama31-70b-sft	Rationale + answer	Llama 3.1 70B Instruct	0	12,515	15,661	28,176	31.1519%
teacher-decomposition-llama31-8b-sft	Question decomposition	Llama 3.1 8B Instruct	3,618	18,122	5,031	26,771	29.5985%
teacher-decomposition-llama31-70b-sft	Question decomposition	Llama 3.1 70B Instruct	0	56,814	15,661	72,475	80.1298%

Coverage uses the 90,447-example canonical HotpotQA distractor/train source as the denominator. Missing archival annotations are not filled with templates, imputed text, or newly generated outputs.

Which configuration should I use?

• Use a teacher-cot-* configuration to train an evidence-grounded reader whose assistant target begins with the short answer and continues with the recovered rationale.
• Use a teacher-decomposition-* configuration to train a model to turn a multi-hop question into simpler intermediate questions.
• Select 8b or 70b according to the reported teacher label and the coverage required by the experiment. Do not silently concatenate configurations unless repeated source questions across teachers are intended.

Row schema

All four configurations use the same schema.

Field	Type	Description
source_id	string	Official HotpotQA example ID and canonical cross-repository join key
source_row	int64	Verified zero-based position in the pinned canonical training split
annotation_id	string	Stable digest-derived identity for the resolved annotation
task	string	cot_reader or question_decomposer
teacher_model_id	string	Model identity reported by the archived experiment artifacts
teacher_model_revision	string	not-recorded because the historical immutable endpoint checkpoint was not retained
prompt_id	string	Identifier of the documented generation protocol
question	string	Original HotpotQA question
answer	string	Original short reference answer
question_type	string	HotpotQA question type
difficulty	string	easy, medium, or hard
gold_context	list of structs	Gold paragraphs with titles and sentence lists
evidence	list of structs	Gold title, sentence index, and resolved sentence text
annotation	string	Formatting-normalized recovered teacher output
messages	list of structs	Ordered system, user, and assistant SFT turns
source_artifact	string	Archived artifact name used for recovery
source_artifact_row	int64	Row position within that artifact
artifact_sha256	string	SHA-256 digest of the source artifact
artifact_variant	string	Recovery condition or shard role
normalization	string	Formatting-only normalization applied to the annotation

An abbreviated rationale row has this shape:

{
  "source_id": "5a7a06935542990198eaf050",
  "source_row": 0,
  "annotation_id": "adcb9583...",
  "task": "cot_reader",
  "teacher_model_id": "meta-llama/Meta-Llama-3.1-70B-Instruct",
  "teacher_model_revision": "not-recorded",
  "prompt_id": "teacher-cot-v1",
  "question": "Which magazine was started first ...?",
  "answer": "Arthur's Magazine",
  "question_type": "comparison",
  "difficulty": "medium",
  "gold_context": [
    {"title": "Arthur's Magazine", "sentences": ["..."]}
  ],
  "evidence": [
    {"title": "Arthur's Magazine", "sentence_index": 0, "sentence": "..."}
  ],
  "annotation": "... recovered teacher rationale ...",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "answer: ***Arthur's Magazine***\nrationale:\n..."}
  ],
  "source_artifact": "COT/CoT_llama70_medium_1.csv",
  "source_artifact_row": 0,
  "artifact_sha256": "829ec0c5...",
  "artifact_variant": "medium-positional-shard",
  "normalization": "Unicode NFC; LF line endings; trailing-space and edge trim"
}

The example is abbreviated for readability. Load the Parquet row for the complete text and nested fields.

Construction and provenance

The release was constructed as a source-preserving recovery pipeline:

1. Pin the official HotpotQA distractor/train source and the canonical Bactrainus representation.
2. Retain complete, non-empty archived teacher records and map their positional rows to canonical source_row values.
3. Verify source_row, source_id, gold context, and evidence against the canonical record.
4. Apply formatting-only normalization: Unicode NFC, LF line endings, trailing-whitespace removal, bounded blank lines, and outer trim.
5. Resolve overlaps deterministically. When a complete rerun and a base record map to the same task, teacher label, and source_id, the documented rerun takes precedence. Only the resolved output is published; an older duplicate version is not retained as a second row.
6. Attach the normalized annotation and an ordered three-turn SFT conversation, then write immutable Parquet shards.

No spelling correction, paraphrasing, factual editing, or synthetic completion is applied to recovered annotations. See DATA_PROVENANCE.md for the precise source-linkage and conflict-resolution policy.

Integrity checks

Release validation enforces:

• exactly 198,660 rows across four configurations and 42 Parquet shards;
• the expected per-configuration row and difficulty counts shown above;
• unique source_id and annotation_id values within every configuration;
• exact canonical source_row to source_id alignment;
• non-empty annotations and the expected task and teacher fields;
• ordered system, user, and assistant message roles;
• title, sentence-index, and sentence-text agreement for every evidence item; and
• source-artifact digest membership and exclusion of evaluation/result artifacts.

Reported teacher identity

The fields teacher_model_id preserve the identities reported by the archived experiment artifacts and manuscript:

• meta-llama/Meta-Llama-3.1-8B-Instruct
• meta-llama/Meta-Llama-3.1-70B-Instruct

The historical serving logs and immutable endpoint checkpoint revisions were not retained. Therefore, these are reported teacher labels, not cryptographic attestations of the exact served weights, and every row states teacher_model_revision: "not-recorded". The release does not invent a revision hash or imply a stronger provenance claim than the surviving records support.

Limitations and responsible use

• Teacher annotations are machine-generated supervision, not human-authored explanations or verified proofs. They can contain unsupported steps, omissions, stylistic artifacts, answer leakage, or factual errors even when the short reference answer is correct.
• The four configurations have different, non-random coverage conditions. They should not be treated as interchangeable representative samples of all HotpotQA difficulty levels.
• gold_context, evidence, and answer are supervision fields. Using them at inference time would create oracle access that is unavailable in a standard open-context QA setting.
• A rationale can be useful as a training target without faithfully representing a model's internal computation. Do not interpret it as mechanistic evidence.
• HotpotQA is derived from English Wikipedia and may reproduce errors, biases, or outdated statements from its source material.
• Before training or deployment, inspect samples, define content filters appropriate to the application, and evaluate the resulting model independently. This dataset contains no evaluation results and makes no performance guarantee.

License and model terms

The dataset release is licensed under Creative Commons Attribution-ShareAlike 4.0 International. Retain upstream HotpotQA and Bactrainus attribution and indicate modifications; adaptations must follow the ShareAlike requirement. See LICENSE and ATTRIBUTION.md.

Built with Llama. The dataset license and the terms governing Meta Llama models address different artifacts. Review the applicable Llama 3.1 Community License, Acceptable Use Policy, and model cards before using the reported teachers or distributing downstream models.

Citation

If this release supports your work, cite Bactrainus and HotpotQA:

@article{barati2025bactrainus,
  title   = {Bactrainus: Optimizing Large Language Models for Multi-hop Complex Question Answering Tasks},
  author  = {Barati, Iman and Ghafouri, Arash and Minaei-Bidgoli, Behrouz},
  journal = {arXiv preprint arXiv:2501.06286},
  year    = {2025},
  url     = {https://arxiv.org/abs/2501.06286}
}

@inproceedings{yang2018hotpotqa,
  title     = {{HotpotQA}: A Dataset for Diverse, Explainable Multi-hop Question Answering},
  author    = {Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D.},
  booktitle = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  year      = {2018},
  pages     = {2369--2380},
  doi       = {10.18653/v1/D18-1259}
}

A dataset-specific citation and the Llama paper citation are provided in ATTRIBUTION.md.

Links

• Canonical Bactrainus HotpotQA dataset
• Bactrainus code repository
• Bactrainus models
• Bactrainus article
• HotpotQA project