YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
geode-internals
Public, token-free companion to mhieuuu/geode-store
for the MARS V project "Mechanistic Understanding of Elicitation vs. Teaching"
(code: https://github.com/Hieuuum/elicit-vs-teach).
geode-store holds every run's final checkpoint + manifest. This repo holds the
heavy per-run internals used by the mechanistic analyses:
runs/<run_id>/model/ final weights (duplicated from geode-store for convenience)
runs/<run_id>/manifest.json run manifest (schema: specs/00-interfaces.md)
runs/<run_id>/snapshots/step_<k>/ target-stage LoRA snapshots (adapter.safetensors, ~48 MB each)
runs/<run_id>/snapshots/base/ the frozen base state the adapters apply to
runs/<run_id>/sft_snapshots/step_<k>/ full-FT parent snapshots (full HF save_pretrained dirs)
runs/<run_id>/logs/ prequential.jsonl, gradstats.jsonl (per-step grad norms)
activations/<run_id>/... activation dumps (geode.zoo.activations store)
results/ analysis outputs (parquet/csv)
Pull a run with snapshots:
python3 experiments/training-run/scripts/hf_checkpoint.py pull \
--run-id evt-ts38mt-pp-n21544 --repo-id mhieuuu/geode-internals --with-snapshots
LoRA checkpoints load only via geode.zoo.load_model (scaling is α/(2r), not PEFT's α/r).
First family here: ts38mt (see EXPERIMENTS.md §6.22 in the code repo).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support