---
license: llama3
language:
  - en
library_name: transformers
pipeline_tag: text-generation
base_model: meta-llama/Meta-Llama-3-70B-Instruct
base_model_relation: finetune
datasets:
  - bactrianus/bactrainus-hotpotqa
tags:
  - bactrainus
  - bactrianus
  - llama-3
  - hotpotqa
  - multi-hop-qa
  - answer-generation
  - legacy
---

# Bactrainus HotpotQA Reader — Llama 3 70B Instruct

<p align="center">
  <img src="assets/llama70.png" alt="Bactrainus Llama 3 70B reader artwork" width="720">
</p>

## Artifact identity

- **Status:** complete merged causal-language-model checkpoint
- **Base model:** `meta-llama/Meta-Llama-3-70B-Instruct`
- **Audited Hub revision:** `8f76e68c65955843fffbf5c9b0e0ae1e446fbf1a`
- **Public artifact date:** August 2024
- **Role:** large direct answer reader over supplied evidence

This is a historical **Llama 3** artifact. The revised Bactrainus manuscript uses **Llama 3.1 70B Instruct** for its controlled large-reader experiments. The two checkpoint identities are not interchangeable.

## Model summary

The model is the large historical Bactrainus reader. It generates a concise answer from a question and evidence selected by an upstream component. It does not select passages or search an external corpus.

## Intended use

- Research on capacity effects in evidence-conditioned multi-hop answer generation.
- Reader-stage diagnostics with validated HotpotQA supporting facts.
- Final-stage generation in a fixed-candidate modular pipeline.

### Out-of-scope use

- Open-domain retrieval or web search.
- Unverified factual generation without supplied evidence.
- Safety-critical or high-stakes applications.
- Treating revised Llama 3.1 results as measurements of this Llama 3 checkpoint.

## Input and output contract

Input should contain one question and a compact, title-preserving serialization of selected evidence. Output is free-form text intended to contain the answer. Parse and normalize the output explicitly.

The exact historical prompt serialization is not embedded in the public model configuration. Users seeking strict legacy replication must pin the model revision and preserve their complete prompt/runtime manifest.

## Loading

The 70B checkpoint requires substantial accelerator memory. Pin the revision and select a supported sharding or quantization strategy:

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "bactrianus/HotpotQA-Reader-Llama-3-70B-Instruct"
REVISION = "8f76e68c65955843fffbf5c9b0e0ae1e446fbf1a"

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()
```

The loading example does not prescribe a production serving topology.

## Training data and lineage

The checkpoint is derived from Meta Llama 3 70B Instruct and adapted on HotpotQA-derived reader supervision. Its matching canonical training view is [`reader-sft`](https://huggingface.co/datasets/bactrianus/bactrainus-hotpotqa/tree/v1.0.0/data/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`. Exact byte identity with the historical 2024 files is not asserted.

```python
from datasets import load_dataset

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

The revised manuscript's `reader_70b.yaml` describes a later Llama 3.1 experiment and must not be presented as this historical checkpoint's verified training manifest.

## Evaluation boundary

No evaluation predictions or results are included here. See the [paper](https://arxiv.org/abs/2501.06286) for reported experiments and their artifact boundary. A fresh evaluation of this exact pinned revision is a new measurement.

## Limitations

- The checkpoint is specialized for English HotpotQA-style evidence-conditioned answering.
- It depends on upstream evidence recall and cannot reliably repair missing evidence.
- Full paragraphs and distractors can reduce answer quality despite larger model capacity.
- Serving cost, latency, throughput, and energy were not established by this model card.
- Model generations can contain unsupported statements or formatting errors.
- The historical release does not contain a complete environment/prompt manifest.

## License and attribution

The weights remain subject to the [Meta Llama 3 Community License](https://github.com/meta-llama/llama3/blob/main/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](https://creativecommons.org/licenses/by-sa/4.0/). Bactrainus code is Apache-2.0 licensed.

## Citation

```bibtex
@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}
}
```