Bactrainus HotpotQA Sentence 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:
0ba9d0d84d7913f3d614061ebd71dca8690e9f93 - Public artifact date: August 2024
- Role: supporting-sentence selection within supplied paragraphs
The repository's historical “Retriever” name denotes generation of supporting-fact references within selected candidate paragraphs. It is not corpus retrieval.
This is a legacy Llama 3 artifact, separate from the revised manuscript's Llama 3.1 sentence selector.
Model summary
The model receives a multi-hop question and selected paragraphs, then emits the exact title and zero-based sentence index of each predicted supporting fact. Its output is designed to feed a reader through a compact facts interface.
Intended use
- Supporting-fact selection for English HotpotQA-style inputs.
- Sentence-level interface in a paragraph-to-sentence cascade.
- Analysis of exact set prediction and error propagation.
Out-of-scope use
- Searching outside the supplied paragraphs.
- Treating selected sentences as verified evidence without checking them against the input.
- General sentence ranking in unrelated domains.
- Claiming direct reproduction of revised Llama 3.1 values.
Input and output contract
Input must preserve exact paragraph titles, sentence order, and zero-based indices. Output should be parsed into a deduplicated set of (title, sentence_index) pairs.
Every title and index must be validated against the current input instance. Unknown titles, negative indices, out-of-range indices, or malformed pairs should remain explicit errors rather than being silently repaired. The exact historical prompt is not embedded in the public configuration.
Loading
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_ID = "bactrianus/HotpotQA-Sentence-Retriever-Llama-3-8B-Instruct"
REVISION = "0ba9d0d84d7913f3d614061ebd71dca8690e9f93"
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 supporting-fact annotations. Its matching canonical training view is sentence-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 a clean deterministic view, not a claim of byte identity with all historical files.
from datasets import load_dataset
train = load_dataset(
"bactrianus/bactrainus-hotpotqa",
"sentence-selector-sft",
split="train",
revision="v1.0.0",
)
The revised Llama 3.1 settings in sentence_selector.yaml do not retrospectively define the training history of this legacy weight artifact.
Evaluation boundary
No evaluation artifacts or score claims are included. Use exact HotpotQA supporting-fact EM/F1 with a documented parser and pinned split. See the paper for the system-level analysis.
Limitations
- Performance depends on the paragraph selector's recall.
- Exact set prediction penalizes one missing or extra sentence.
- Generated indices are vulnerable to off-by-one and formatting errors.
- The model assumes HotpotQA-style title and sentence boundaries.
- It was not validated as a general evidence-verification model.
- 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}
}