Source-linked AI summary
SHARE: Social-Humanities AI for Research and Education
João Gonçalves, Sonia de Jager, Petr Knoth, David Pride, Nick Jelicic
TL;DR
SSH research lacks language models and interfaces designed to support its open-ended, critical work rather than merely generate plausible text. The paper develops SSH-pretrained SHARE models and the non-generative MIRROR interface, finding that SHARE-14B performs similarly to Phi-4 with 100 times fewer training tokens while MIRROR highlights unexpectedness for reflection. The authors present this combination as an alternative way to use LLMs in SSH, while acknowledging unresolved risks around shortcut use and inherited bias.
Problem
SSH research needs LLM technologies that accommodate open-ended exploration and critical thinking, but mainstream models and interfaces are not suited to several roles of language in SSH work.
Method
The paper develops SHARE, SSH-pretrained causal language models, and MIRROR, an interface that extracts token probabilities and signals unexpectedness without generating text.
Results
SHARE-14B scores .80 on the SSH Cloze benchmark, close to Phi-4's .82 despite using 100 times fewer training tokens, while MIRROR prototypes reflective review of SSH texts.
Takeaways & Limitations
SHARE and MIRROR demonstrate a domain-specific, non-generative route for using causal LLMs to support reflection on SSH writing and disciplinary expectations.
Takeaways & Limitations
MIRROR may encourage users to treat unexpectedness highlights as shortcuts, and SHARE inherits systemic biases from its SSH training scholarship.
Abstract
from arXiv · showhide
This intermediate technical report introduces the SHARE family of base models and the MIRROR user interface. The SHARE models are the first causal language models fully pretrained by and for the social sciences and humanities (SSH). Their performance in modelling SSH texts is close to that of general purpose models (Phi-4) which use 100 times more tokens, as shown by our custom SSH Cloze benchmark. The MIRROR user interface is designed for reviewing text inputs from the SSH disciplines while preserving critical engagement. By prototyping a generative AI interface that does not generate any text, we propose a way to harness the capabilities of the SHARE models without compromising the integrity of SSH principles and norms.
1 Introduction
LLMs raise concerns for SSH research because chat-oriented, output-focused use may undermine quality, bias awareness, and critical thinking. SHARE and MIRROR address this by developing SSH-specific models and an interface that signals unexpectedness for reflection rather than generating text.
- LLM adoption in scientific writing raises concerns about quality, bias, and erosion of critical thinking, especially in SSH domains.
- Commercial chat interfaces emphasize producing outputs, leaving it unclear whether SSH risks arise from the technology or its deployment.
- SHARE models are transformer-based LLMs pretrained exclusively on content relevant to SSH scholarship.
- MIRROR uses SHARE to signal novelty and discovery through unexpectedness in research outputs rather than simply generating text.
- Together, SHARE and MIRROR support self-reflection on academic writing and expose disciplinary biases and constraints through deviations from expected outputs.
2 Related work
SSH-focused language-model development remains limited compared with work centered on technical fields and general scientific claims. SHARE fills a stated gap by providing large autoregressive models pretrained on SSH-relevant text.
- The project's data is primarily English, which the authors identify as a limitation for SSH fields and a target for future multilingual work.
- AI development and impact claims in scientific work have often centered on mathematics, coding, engineering, and biochemistry rather than SSH domains.
- SSH research poses challenges for techniques relying on verifiable rewards because its interpretative nature is less suited to those evaluation settings.
- Existing SSH-focused efforts include SSciBERT for social-science abstracts and BERT-NLI for text classification of political content.
- SHARE addresses the stated gap in autoregressive SSH models by providing 4B and 14B parameter models pretrained on more than tens of billions of SSH-relevant tokens.
3 The role of expectations
The paper reframes next-token prediction in SSH as a way to make disciplinary expectations visible. MIRROR therefore highlights unexpectedness for reflection, drawing on the idea that deviations can mark either problems or insight.
- Chat interfaces establish an expectation that causal LLMs are primarily useful for generating the most likely continuation of user input.
- SSH scholarship values exploring disconnected concepts, reframing ideas, and questioning structures that shape language and academic work.
- For an SSH-pretrained autoregressive model, next-token prediction makes explicit what the field expects in disciplinary writing.
- SHARE signals deviations from expectation so MIRROR can support disciplinary reflection and exploration instead of overproductive automation.
- The approach assumes that expectation violations can sometimes be preferable to confirmations, while deviations may also indicate undesired aspects.
- Unexpectedness can characterize successful scientific writing because new theoretical lenses and counter-intuitive claims may provide insight.
- MIRROR uses colour-coded labels to signal unexpected tokens in user texts as its default interaction with SHARE models.
4 The SHARE Family of Models
The SHARE family is built from SSH-focused text selected across Wikipedia, Project Gutenberg, PeS2o, and CORE, then used to train decoder-only transformer models. The report describes SHARE-4B training and ongoing SHARE-14B pretraining.
- Dataset construction: The SSH-focused dataset combines Wikipedia articles, Project Gutenberg books, and academic publications from PeS2o and CORE.Academic publications were primarily sourced from PeS2o and CORE, with a smaller publisher- and author-agreed subset.
- Dataset construction: SSH documents are identified using metadata-based heuristics and machine-learning classification, including AllenAI Field of Science labels.For CORE, Field of Science labels are generated from titles, abstracts, and available publication venues.
- Dataset construction: CORE supports the dataset through large-scale acquisition, deduplication, parsing, and metadata enrichment of scholarly content.These processes underpin construction of domain-specific datasets and the empirical analyses.
- Model design: The SHARE models are decoder-only transformers based on Phi-4, with an expanded 8192-token context and a custom 50 000-token BPE tokenizer for SHARE-4B.The initial context length was 4096 tokens, and the tokenizer vocabulary is smaller than Phi-4’s 100 352 tokens.
- Training: SHARE-4B was trained for 2 epochs over 28 billion tokens using 8 NVIDIA-A100 GPUs for 656 hours, while SHARE-14B had reached 96 billion tokens.The planned SHARE-14B run targets 630 billion tokens over 2 epochs; the report describes this as 15% complete.
5 The MIRROR interface
MIRROR provides analytical interaction with pretrained SHARE models rather than conventional text generation. Its design emphasizes explicit SSH expectations, critical engagement, and reduced risks associated with prescriptive or anthropomorphic use.
- Rationale: Because SSH-specific SFT, DPO, and RLHF datasets were unavailable, the project shifted toward an interface for extracting analytical insights without chat dialogue.General SFT experiments were observed to degrade text-generation performance.
- Design principles: MIRROR is guided by interaction with SHARE in its pretrained state, explicit SSH expectations, and respect for publishers, authors, and educators.The principles aim to avoid prescriptive or normative outputs and minimize copyright infringement.
- Interface concept: MIRROR reverses plagiarism detection by using SHARE to estimate how unexpected a user’s text is within domain-specific language.The approach relies on SHARE’s open weights, domain specificity, and lack of further alignment to extract token probabilities.
- Visualizations: The interface visualizes token-level unexpectedness through surprisal and entropy, including heatmaps, ranked alternatives, and sentence- or paragraph-level aggregation.Hovering over a token presents expected continuations based on the preceding text rather than ordinary thesaurus similarity.
6 Results
The results evaluate SHARE through domain-specialization comparisons, an SSH-specific Cloze benchmark, and qualitative MIRROR use cases. SHARE shows stronger SSH specialization and competitive Cloze performance, while MIRROR highlights both errors and potentially innovative deviations without generating text.
- 6.1 Relative specialization per scientific field: SHARE fits SSH fields relatively better than STEM fields compared with Phi-4, indicating a stronger specialization skew toward SSH domains.The comparison uses log perplexity gaps across scientific domains; lower values indicate better relative SHARE fit.
- 6.2 The SSH Cloze Benchmark: The initial SSH Cloze benchmark samples 275 out-of-distribution abstracts across 11 Web of Science fields and is intended to emphasize SSH familiarity.The benchmark uses recent abstracts and equivalent-token predictions, but its initial version is planned for expansion.
- 6.2 The SSH Cloze Benchmark: SHARE-14B reaches .80 on the SSH Cloze benchmark, close to Phi-4’s .82 despite using 100 times fewer training tokens.It also outperforms Pythia-12B (.62) and Olmo-2-13B (.74), while SHARE-4B marginally exceeds Pythia-3B (.64) at .66.
- 6.3 Qualitative insights: MIRROR’s anecdotal tests show both SHARE models can flag typos and stylistic deviations, while SHARE-14B identifies more nuanced unexpected choices.The qualitative evaluation covers negative expectation violations such as mistakes and positive violations such as innovation and challenges to dominant narratives.
- 6.3 Qualitative insights: MIRROR can surface factual inconsistencies, missing expected terms, and unexpected concepts while leaving authors to decide whether revisions or expectation-conforming terms are warranted.The interface signals unexpected tokens rather than directly inserting them, supporting author reflection on factual accuracy, omissions, and novelty.
7 Discussion
SHARE and MIRROR address a gap between SSH research needs and mainstream LLM deployment by combining SSH-focused models with an interface designed to support reflection. The report presents domain-specific models as an alternative to commercial extraction while acknowledging that responsible use and impact remain unsettled.
- SHARE identifies a mismatch between mainstream LLMs and the roles language plays in SSH research, proposing artifacts developed by, with, and for SSH scholars.
- Quantitative results show SHARE is more aligned with SSH domains than STEM, while curated data makes it a reasonable proxy for open-access SSH writing.
- MIRROR reframes bias, creativity, and critical thinking as characteristics of language artifacts that SSH fields should examine rather than merely correct through model alignment.
- The report cannot yet assess SHARE and MIRROR’s most critical contribution: whether they will shift debate toward SSH shaping AI models and interfaces.
8 Compliance and ethics
The report treats SHARE’s ethical implications across data, training, deployment, and inference, linking model development to interface design. It reports limited early memorization and safety concerns but identifies unresolved risks involving bias and non-reflective use.
- SHARE’s research-and-education purpose requires ethical standards and compliance checks throughout data collection, training, evaluation, and deployment.
- Data: Open-access availability does not permit reproducing or commercially exploiting academic works without proper attribution and attention to non-commercial labels.
- Data: SHARE uses EU text-and-data-mining exceptions and a Responsible AI License that forbids commercial use, model distillation, and unrestricted text-generation applications.
- Training: The project uses scaling laws and computing-efficiency measures to determine training resources and reduce the environmental burden of model development.
- Inference: MIRROR extracts token probabilities from user-written text to support self-reflection and critical thinking rather than replacing student work or research argument development.
- Inference: MIRROR makes inherited SSH biases explicit but does not resolve them, and downstream chat systems may present biased outputs with less transparency.
- Inference: Early deterministic tests found minimal memorization risk, with replication limited to disclaimers or standard headers rather than copyright-protected content.
- Inference: Early supervised instruction-tuning experiments also suggest safety concerns are unlikely to be pressing because many hazardous domains were absent from the curated SSH training data.
9 Conclusion
SHARE and MIRROR respond to concerns about critical-thinking erosion, extractive research-data use, and STEM dominance in general-purpose models. They offer domain-specific models and interfaces designed for critical self-reflection in SSH work.
- SHARE and MIRROR offer responsible, domain-specific models and interfaces attuned to the critical self-reflection required in SSH research and education.
11 Author contributions
The author-contribution statement assigns responsibilities across project leadership, planning, training, data work, conceptualization, benchmarking, interface development, writing, and dataset facilitation.
- João Gonçalves led the project and contributed to model training, benchmarking, interface development, and manuscript writing.
- Sonia de Jager contributed project planning, data collection, philosophical conceptualization and analysis, and final manuscript writing.
- Nick Jelicic handled data preparation and filtering, technical consultancy, data-section writing, and project planning.
- Petr Knoth and David Pride contributed CORE dataset description and facilitation, final revisions, and project planning.