Skip to content
View mac-n's full-sized avatar
0 followers · 1 following

Block or report mac-n

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
mac-n/README.md

Niamh McCombe

Independent AI researcher, Northern Ireland. PhD in machine learning (Ulster University, Intelligent Systems Research Centre). Website: glassnest.ai

Current work

Reward hacking is largely a property of the scaffold · write-up Same model, same 103 unsatisfiable coding tasks, same prompt: 63% cheating inside ImpossibleBench's own agent scaffold, 2% inside an ordinary coding agent. Persistence after a failed submission is the trigger. Data, harness and analysis scripts in the repo; every number recomputes from the CSVs.

Embodied cognition in transformers · findings as slides · full write-up Lakoff image schemas (UP-DOWN, BALANCE, FORCE…) recovered from single-token activations in Pythia, GPT-2 and Llama. Steering along UP makes the model happier. Inflectional morphology organises itself around the schema axes in the transformer but not in GloVe or word2vec, so the schema system is learned, not copied from word statistics. BALANCE is coupled to the model's own computation: residual norm in Pythia, attention entropy in GPT-2 and Llama. Every script, raw output, pre-registration and lab notebook is in the repo, including the failures.

GLAS A transparent-by-construction language model: a tree of specialists competing on prediction error instead of an attention stack. Beats a parameter-matched transformer at 100K and 1M parameters on character-level Shakespeare. Working on how to scale it.

Earlier

Applied machine learning for dementia diagnosis (missing-data imputation, graph neural networks, assessment optimisation): Google Scholar.

Pinned Loading

  1. lakoff-schemas-in-transformers lakoff-schemas-in-transformers Public

    Embodied cognition in transformers: Lakoff image schemas in Pythia, GPT-2 and Llama. UP steers happiness; the schema system is learned, not inherited from word statistics; BALANCE couples to residu…

    Python 1

  2. scaffold-effect scaffold-effect Public

    Reward hacking is largely a property of the scaffold: same model, same 103 tasks, 63% cheating in ImpossibleBench's harness vs 2% in an ordinary coding agent

    Python

  3. Clustering-GNN Clustering-GNN Public

    Code for 'Alzheimer’s Disease Classification Using Cluster-based Labelling for Graph Neural Network on Tau PET Imaging and Heterogeneous Data '

    Jupyter Notebook 12 1

  4. mac-n.github.io mac-n.github.io Public

    Research blog. Embodied Cognition in Transformers: image schemas, metaphorical mappings, and substrate coupling in language models trained on text alone.

    HTML 1

  5. BHI-missing-data BHI-missing-data Public

    code for BHI missing data paper

    R 1 2

  6. DLB DLB Public

    Code for DLB diagnosis problem

    R 1