cs.CLApr 13, 2026
SaveSHARE: Social-Humanities AI for Research and Education
Organizations: Erasmus University Rotterdam, NL · Open University, UK
Abstract
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.
Figures & tables
Figure 1: Two pipelines for LLM use in scholarly work. The standard pipeline generates expected text from a general-purpose aligned model; the SHARE + MIRROR pipeline uses an SSH-specialised base model to signal unexpected tokens for the reader to reflect.
| Main topic | Notes |
|---|---|
| Business | |
| Communication | Excluding animal and plant communication |
| Culture | |
| Economy | |
| Education | |
| Geography |
Table 1: Main Wikipedia Topics
| Category | Name |
|---|---|
| Class B | Philosophy, Psychology, Religion |
| Class C | History |
| Class D | History: General and Eastern Hemisphere |
| Class G | Geography, Anthropology, Recreation |
| Class H | Social sciences |
| Class J | Political science |
Table 2: Gutenberg Categories
| Field of Science |
|---|
| Art |
| Business |
| Economics |
| Geography |
| Education |
| History |
Table 3: Fields of Science Categories
Figure 2: The figure shows the average difference in log perplexity between the SHARE and Phi-4 models per scientific domain following the FoS classifier. Lower values mean that SHARE fits better to that scientific domain in relation to Phi-4. Error bars indicate 95% confidence intervals.
Figure 3: The figure shows the average difference in log perplexity between the SHARE and Phi-4 models based on the faculty that the author is affiliated to at Erasmus University Rotterdam. Lower values mean that SHARE fits better to research outputs of that faculty in relation to Phi-4. Error bars indicate 95% confidence intervals.
| Model | Size | Training tokens | Raw accuracy | Prior corrected |
|---|---|---|---|---|
| Phi-4 | 14B | 9.8T | 81.8% | 81.8% |
| SHARE | 14B | 96B | 77.1% | 79.6% |
| OLMO-2 | 7B | 4T | 78.2% | 76.4% |
| OLMO-2-Step-20k | 13B | 168B | 74.9% | 73.8% |
| Phi-4 | 4B | 5T | 73.8% | 69.8% |
| SHARE | 4B | 28B | 69.8% | 66.2% |
Table 4: SSH Cloze LLM Benchmark Results. Token counts with asterisk * denote that the token count was estimated from word counts. Models are ranked by prior corrected accuracy.
Figure 4: Prior corrected accuracy on the SSH Cloze benchmark (vertical axis) based on estimated flops (horizontal axis, logarithmic scale). Data points in the top left quadrant of the chart indicate more efficient models, achieving better performance with lower training compute.
Figure 5: Example use of SHARE-MIRROR for typo and style detection, where tokens highlighted in red show deviations from model expectations. The top output refers to the SHARE-4B model while the bottom output refers to SHARE-14B.
Figure 6: Three tests of reactions to factual inaccuracies in text by the SHARE models. The top and middle examples show, respectively, unexpectedness signalled by the SHARE-4B and the SHARE-14B models, with token predictions following "proposed by" to the left. The bottom example modifies the text by switching the word "concept" by "fact" and shows uncertainty for the SHARE-14B model.
Figure 7: The figure illustrates how SHARE-4B (top) and SHARE-14B (bottom) can be leveraged to identify unexpected tokens in a discussion section. Unlike examples in previous figures, these deviations from expectation, highlighted in red, likely signal positive deviations from expectations rather than factual or grammar mistakes.
Figure 8: Analysis of unexpectedness of the first paragraph of Gregory Gondwe’s response article to the presidential address of the International Communication Association 2025 Conference. The tokens in red show how research framed from African locations is signalled as unexpected by a model trained on English open access SSH scholarship.
Figure 9: The figure above signals in green which tokens of a text Sonn and Hsu (2022) used in model training were correctly predicted (memorized) by the model while red tokens show instances where model generation would deviate from the original text. The high prevalence of red tokens shows that the model is not memorizing copyrighted works.
More in this topic
Language Modeling
Language modeling develops systems that learn patterns in text and other sequences to understand, predict, and generate language.
See all 519 papers