HotpotQA-Reader-Llama-3-8B-Instruct

Bactrainus HotpotQA Reader — Llama 3 8B Instruct

<p align="center"> <img src="assets/models.png" alt="Bactrainus Llama 3 model collection artwork" width="720"> </p>

Artifact identity

• Status: complete merged causal-language-model checkpoint
• Base model: meta-llama/Meta-Llama-3-8B-Instruct
• Audited Hub revision: 98c63afd55b5b4bd46890165e85e69c7d7d10a3d
• Public artifact date: August 2024
• Role: direct answer reader over supplied evidence

This is a historical Llama 3 artifact. The revised Bactrainus manuscript reports controlled architecture experiments with Llama 3.1 checkpoints. This card does not relabel, rebase, or claim that these weights are the revised manuscript checkpoint.

Model summary

The model is a HotpotQA-focused reader. It receives a multi-hop question together with already selected evidence and generates a concise answer. Evidence selection is outside this checkpoint's responsibility.

This model is appropriate for studying the reader stage independently or as the final component of a fixed-candidate selector--reader pipeline.

Intended use

• Research on answer generation from compact multi-document evidence.
• Reader-stage ablations in the English HotpotQA distractor setting.
• Integration behind a paragraph/sentence selector that emits validated evidence.

Out-of-scope use

• Open-domain retrieval or searching Wikipedia.
• Treating the model as a source of verified facts without supplied evidence.
• Safety-critical, legal, medical, or high-stakes decision support.
• Claiming direct reproduction of revised Llama 3.1 paper values.

Input and output contract

Input should contain:

1. one question;
2. a deterministic serialization of selected supporting sentences, including their titles;
3. an instruction to return a concise answer.

Output is free-form text and must be normalized and parsed before evaluation. The public legacy configuration does not preserve a complete prompt-version manifest; an illustrative prompt should not be treated as a byte-exact reconstruction of the 2024 training serialization.

Loading

Pin the immutable revision:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "bactrianus/HotpotQA-Reader-Llama-3-8B-Instruct"
REVISION = "98c63afd55b5b4bd46890165e85e69c7d7d10a3d"

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()

Review the Meta Llama 3 license and choose hardware, precision, and generation limits appropriate for your environment before running inference.

Training data and lineage

The checkpoint is derived from Meta Llama 3 8B Instruct and adapted for HotpotQA-derived reader supervision. HotpotQA supplies English Wikipedia questions, candidate paragraphs, answers, and supporting-fact annotations.

The matching canonical training view is reader-sft, pinned to dataset tag v1.0.0. It contains all 90,447 training source IDs and remains joinable to every other view through source_id; byte-for-byte identity with every historical 2024 training file is not asserted.

from datasets import load_dataset

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

Do not project the revised Llama 3.1 optimizer and LoRA configuration onto this historical merged checkpoint unless an independently verified run manifest establishes that identity.

Evaluation boundary

No predictions, evaluation results, or leaderboard claims are bundled with this model card. For the methodology and the distinction between historical artifacts and revised experiments, see the Bactrainus paper and the clean code repository.

Limitations

• The model was specialized for English HotpotQA-style inputs and was not validated as a general reader.
• Accuracy depends strongly on evidence quality; missing facts cannot be recovered reliably downstream.
• Extra paragraphs can introduce substantial context noise.
• Generated answers may be unsupported, malformed, or more verbose than the expected short answer.
• Wikipedia-derived data inherits temporal, cultural, and coverage biases.
• The historical release lacks a complete immutable environment and prompt manifest.

License and attribution

The weights remain subject to the Meta Llama 3 Community License and Acceptable Use Policy. Redistribution must include the upstream agreement and required notices.

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. The Bactrainus code is separately licensed under Apache-2.0.

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}
}