You can deploy this model and start earning money today!

mhieuuu/
geode-internals

Name: geode-internals
Availability: InStock

safetensors

Views

No views yet

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:

bash

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).