The rise of large language models informed by not so large corpora of training data
Bibliographic record
Abstract
CSA ResearchThe general public first became aware of generative AI (GenAI) on November 15, 2022 when Meta launched its Galactica large language model (LLMs).Claims from its creators that it could "reason" about scientific knowledge proved to be overly optimistic and public social media posting about its limitations led Meta to withdraw the model just three days later, on November 18 (Dickson 2022).At the time, Meta's chief AI scientist, Yan LeCun, publicly blamed the demise of Galactica on public "casual misuse" of the model on X (formerly Twitter) (@ylecun, November 18, 2022).Nevertheless, media coverage introduced the term "hallucination" to a broader audience.With what can now be seen as very fortuitous timing, OpenAI launched its own ChatGPT service less than two weeks after Galactica shut down, filling the void left behind by Galactica for a public eager to try out the seemingly magic power of LLMs.The OpenAI team had learned how not to describe these models from watching Meta's struggles, and ChatGPT became a smash hit, far exceeding its creators' expectations (Heaven 2023). "Magical" English performance may not translateEarly adopters of ChatGPT quickly realized that it produced surprisingly good translations for many languages, leading to discussion in translation and localization circles about whether LLMs would soon replace Neural Machine Translation (NMT).GenAI hype led some companies to lay their localization teams off because they felt they could simply have ChatGPT translate all of their content.At CSA Research, our extensive testing found in April 2023 that although GenAI had significant promise and advantages over NMT in translation, it also had significant drawbacks that kept it from being a viable large-scale approach for enterprise translation (Lommel 2023).
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.016 | 0.076 |
| Meta-epidemiology (narrow) | 0.002 | 0.002 |
| Meta-epidemiology (broad) | 0.003 | 0.002 |
| Bibliometrics | 0.003 | 0.005 |
| Science and technology studies | 0.001 | 0.003 |
| Scholarly communication | 0.005 | 0.014 |
| Open science | 0.003 | 0.005 |
| Research integrity | 0.003 | 0.009 |
| Insufficient payload (model declined to judge) | 0.006 | 0.006 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".