Bactrainus HotpotQA Paragraph Selector — Llama 3 8B Instruct
Artifact identity
- Status: complete merged causal-language-model checkpoint
- Base model:
meta-llama/Meta-Llama-3-8B-Instruct - Audited Hub revision:
1bc650152552e5a0ee7ee4afdfae01abc3bdc76e - Public artifact date: August 2024
- Role: relevant-paragraph selection from the supplied candidate set
The historical repository uses the word “Retriever,” but the model does not query an external corpus. It selects from the fixed candidate set supplied by HotpotQA.
This is a legacy Llama 3 checkpoint. The revised manuscript's paragraph selector is a distinct Llama 3.1 experiment.
Model summary
The paragraph selector receives a question and the candidate paragraphs supplied with that example, then generates the titles of the paragraphs predicted to contain the evidence needed for the answer. Its output is an intermediate interface for a sentence selector or reader.
Intended use
- Fixed-candidate paragraph relevance selection for HotpotQA-style research.
- Upstream component in a modular selector--reader pipeline.
- Analysis of paragraph recall and paragraph-to-sentence error propagation.
Out-of-scope use
- Open-domain retrieval, embedding search, or document indexing.
- General reranking without task-specific validation.
- Safety-critical or production search decisions.
- Using revised Llama 3.1 selector results as measurements of this artifact.
Input and output contract
Input must preserve the question, candidate order, and exact paragraph titles. Output should be parsed as an ordered, deduplicated list of exact candidate titles.
Reject unknown titles. Fuzzy matching can hide model errors and change paragraph-level metrics. The public legacy configuration does not preserve a complete prompt template, so new serialization should be versioned as a new evaluation condition.
Loading
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_ID = "bactrianus/HotpotQA-Paragraph-Retriever-Llama-3-8B-Instruct"
REVISION = "1bc650152552e5a0ee7ee4afdfae01abc3bdc76e"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, revision=REVISION)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
revision=REVISION,
torch_dtype=torch.bfloat16,
device_map="auto",
)
model.eval()
Training data and lineage
The checkpoint derives from Meta Llama 3 8B Instruct and HotpotQA-derived paragraph-selection supervision. Its matching canonical training view is paragraph-selector-sft, pinned to dataset tag v1.0.0. The view contains all 90,447 training source IDs and remains joinable to every other view through source_id. It is not asserted to reproduce every historical byte of the 2024 training files.
from datasets import load_dataset
train = load_dataset(
"bactrianus/bactrainus-hotpotqa",
"paragraph-selector-sft",
split="train",
revision="v1.0.0",
)
The revised paragraph_selector.yaml is a Llama 3.1 configuration and must not be attached to these weights as a verified legacy run manifest.
Evaluation boundary
This card includes no predictions or results. Evaluate exact title sets against the same pinned HotpotQA split and parsing policy. See the paper for the broader selector--reader analysis.
Limitations
- The model assumes exactly the benchmark-provided candidate boundary.
- A selected paragraph can still contain many irrelevant sentences.
- Title formatting and parser policy can affect exact-match scores.
- Missing a relevant paragraph constrains all downstream stages.
- Wikipedia-derived candidates inherit temporal and coverage bias.
- Historical prompt and environment metadata are incomplete.
License and attribution
The weights remain subject to the Meta Llama 3 Community License and Acceptable Use Policy.
Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc. All Rights Reserved.
Built with Meta Llama 3.
HotpotQA-derived data is licensed under CC BY-SA 4.0. Bactrainus code is Apache-2.0 licensed.
Citation
@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},
doi = {10.48550/arXiv.2501.06286},
url = {https://arxiv.org/abs/2501.06286}
}