YirongSun/LLaSO-Align

LLaSO: A Foundational Framework for Reproducible Research in Large Language and Speech Model

This repository contains the datasets (LLaSO-Align, LLaSO-Instruct, LLaSO-Eval) from the paper LLaSO: A Foundational Framework for Reproducible Research in Large Language and Speech Model.

Fully open corpus + benchmark + reference model for compositional speech-language understanding.

Paper • Code

TL;DR. 25.5M training samples, 20 tasks, 3 modality configurations; 15,044-sample stratified benchmark; 3.8B open reference model - an open-source LSLM stack.

Abstract

The abstract of the paper is the following:

The development of Large Speech-Language Models (LSLMs) has been slowed by fragmented architectures and a lack of transparency, hindering the systematic comparison and reproducibility of research. Unlike in the vision-language domain, the LSLM field suffers from the common practice of releasing model weights without their corresponding training data and configurations. To address these critical gaps, we introduce LLaSO, the first fully open, end-to-end framework for large-scale speech-language modeling. LLaSO provides the community with three essential resources: (1) LLaSO-Align, a 12M-instance speech-text alignment corpus; (2) LLaSO-Instruct, a 13.5M-instance multi-task instruction-tuning dataset; and (3) LLaSO-Eval, a reproducible benchmark for standardized evaluation. To validate our framework, we build and release LLaSO-Base, a 3.8B-parameter reference model trained exclusively on our public data. It achieves a normalized score of 0.72, establishing a strong, reproducible baseline that surpasses comparable models. Our analysis reveals that while broader training coverage enhances performance, significant generalization gaps persist on unseen tasks, particularly in pure audio scenarios. By releasing the complete stack of data, benchmarks, and models, LLaSO establishes a foundational open standard to unify research efforts and accelerate community-driven progress in LSLMs. We release the code, dataset, pretrained models, and results in this https URL .

🔍 What is LLaSO?

LLaSO is the first fully open, end-to-end stack for large-scale speech–language modeling, unifying data, evaluation, and modeling in one framework.

• LLaSO-Align (12.0M): ASR-based alignment for grounding speech in textual semantic space.
• LLaSO-Instruct (13.5M / 20 tasks / 3 modality configs): Multi-task instruction tuning across linguistic, semantic, and paralinguistic objectives.
• LLaSO-Eval (15,044): Stratified benchmark for instruction-following and cross-modality generalization.
• LLaSO-Base (3.8B): Two-stage trained reference model, adapted from LLaVA-style architectures for robust compositional understanding.

✨ Key Features

• Fully Open, End-to-End Stack: Unified release of corpus, benchmark, and model-enabling open-source research and fair comparison in speech-language modeling.
• 25.5M Samples, 20 Tasks, 3 Modality Configurations: Supports all major text ↔ audio combinations (text + audio, audio + text, pure audio), covering linguistic, semantic, and paralinguistic tasks.
• Stratified Evaluation (15,044): Cohesive design between training and test sets enables systematic assessment of instruction following, cross-modality generalization, abstention rate, and stability.
• Robust Reference Model (3.8B): Two-stage training (ASR alignment → instruction tuning), easily reproducible and extensible for further research.
• Empirical Insights: Broader task and modality coverage consistently leads to stronger overall performance, but unseen modality/task configurations (especially pure audio) remain challenging; interleaving and parallel decoding strategies can bridge some gaps.

📂 Data Preparation

The training and alignment scripts expect a single JSON file as input. However, for flexibility, we release the instruction data in multiple subdirectories (e.g., audio_text/, pure_audio/, text_audio/).

• This split format allows users to explore or train with individual modality subsets.
• For full training, these subsets need to be merged into one JSON.

We provide ./llaso/scripts/data_merge.py for this purpose. Use it to combine the JSON files under each modality subdirectory into a single training file.

Dataset Availability

• ✅ LLaSO-Eval is already available on Hugging Face.
• ⏳ LLaSO-Instruct is currently being uploaded (large dataset, uploading in parallel threads; this may take some time due to network limits).
• 📌 LLaSO-Align will be uploaded next.

🗃️ Data Cards

LLaSO Corpus Overview

• Composition: 25.5M samples (12.0M Align + 13.5M Instruct) covering 20 tasks across all major modality configurations (text instr. with audio input, pure audio, audio instr. with text input).
• Overall Task Distribution: 52% linguistic, 8% semantic, 40% paralinguistic.
• Real vs. Synthetic: 71% real-world audio, 29% synthetic speech.
• Design Motivation:
  • Linguistic (ASR) remains foundational for speech–text alignment and generalization.
  • Semantic tasks are intentionally underweighted, as their challenge lies more in language modeling than in speech understanding.
  • Paralinguistic tasks (speaker, accent, emotion, pronunciation scoring) are prioritized to address their underrepresentation in open datasets.
• Flexible Modality Roles: Both audio and text serve as input/instruction, enabling rich compositional interaction patterns.

---

LLaSO-Align (12.0M)

• Goal: ASR-based alignment; encoder & LLM frozen, projector trained for speech-to-text semantic grounding.
• Domains: Conversational, narrative, audiobook, accented speech.
• Templates: 18 instruction types for ASR; unified JSON format for integration.

LLaSO-Instruct (13.5M / 20 tasks)

• Purpose: Multi-task instruction tuning for robust, compositional understanding.
• Task Types: Spans linguistic, semantic, and paralinguistic objectives with a mix of closed- and open-ended formats.
• Modality Configurations:
  • Text instruction + Audio input: X_query^(t,a)
  • Audio instruction + Text input: X_query^(a,t)
  • Pure audio: X_query^(a)
• Label Granularity: Multi-granularity (e.g., coarse→fine age, accent).

LLaSO-Eval (15,044)

• Benchmarking: Strictly stratified; consistent with training data.
• Coverage: All tasks and modality combinations.
• Metrics: Supports abstention rate analysis and cross-modality generalization evaluation.

🚀 Quick Start (Sample Usage)

For detailed installation instructions and how to prepare model weights, please refer to the LLaSO GitHub repository.

Training

Train LLaSO-Base from scratch with:

# Stage 1: Speech-Text Alignment (ASR)
bash llaso/scripts/align.sh

# Stage 2: Multi-task Instruction Tuning
bash llaso/scripts/finetune.sh

Inference

Run model inference on your own data or evaluation set:

python llaso/evaluation/model_eval.py \
  --audio_tower ./whisper-large-v3 \
  --model_path ./LLaSO-Base-3.8B-Instruct \
  --data_path ./LLaSO-Eval/your_eval.json \
  --output_dir ./your_output_dir

📑 How to Cite

If you use LLaSO in your research or applications, please cite our paper:

@misc{sun2025llaso,
      title={LLaSO: A Foundational Framework for Reproducible Research in Large Language and Speech Model},
      author={Yirong Sun and Yizhong Geng and Peidong Wei and Yanjun Chen and Jinghan Yang and Rongfei Chen and Wei Zhang and Xiaoyu Shen},
      year={2025},
      eprint={2508.15418},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.15418},
}