d3banjan/lazy-rudder-checkpoints

Lazy Rudder — adapter checkpoints

Companion artefacts for The Lazy Rudder: How LLMs Absorb Alignment — Axiomatic Bounds on LoRA Geometry (Basu, 2026).

Paper / code: https://github.com/d3banjan/lazy-rudder-paper

What's in here

Path	Base model	Objective	Rank	Final step	Notes
_leak_70m/v2/checkpoints/checkpoint-800/	pythia-70m	DPO	128	800	petri-dish sweep
_leak_160m/v2/checkpoints/checkpoint-800/	pythia-160m	DPO	128	800	petri-dish sweep
_leak/checkpoints/checkpoint-800/	pythia-410m	DPO	16	800	r=16 ablation (v1)
_leak/v2/checkpoints/checkpoint-800/	pythia-410m	DPO	128	800	main-text Pythia-410M DPO
_leak/v3/checkpoints/checkpoint-800/	pythia-410m	CLM	128	800	main-text Pythia-410M CLM
_leak_1b/v2/checkpoints/checkpoint-800/	pythia-1b	DPO	128	800	seed 42
_leak_1b/v3/checkpoints/checkpoint-800/	pythia-1b	CLM	128	800	seed 42
_leak_1b_seed117/v2/checkpoints/checkpoint-800/	pythia-1b	DPO	128	800	seed 117 (data-shuffle bug variant)
_leak_1b_seed117/v3/checkpoints/checkpoint-800/	pythia-1b	DPO	128	800	seed 117 (clean)
_leak_1b_seed117/v4/checkpoints/checkpoint-800/	pythia-1b	CLM	128	800	seed 117
bitfit_dpo_strike/checkpoints/checkpoint-800/	pythia-410m	DPO bias-only	—	800	BitFit
_leak/v2/channel_partition.json	—	—	—	—	precomputed channel partition for sectional autopsy

Each directory contains adapter_config.json + adapter_model.safetensors (LoRA dirs) or config.json + model.safetensors (BitFit), plus trainer_state.json for loss-trajectory plots. Optimizer state and RNG state are intentionally omitted — these checkpoints are for analysis, not training restart.

Pull all of it

pip install huggingface_hub
python -c "from huggingface_hub import snapshot_download; \
  snapshot_download('d3banjan/lazy-rudder-checkpoints', \
                    local_dir='./trained-model-battery/results')"

Or use the shipped helper from the paper repo:

git clone https://github.com/d3banjan/lazy-rudder-paper
cd lazy-rudder-paper
python scripts/fetch_checkpoints.py
make analysis      # regenerate every JSON in results/
make paper         # rebuild manuscript/main.pdf

License

Apache 2.0. Base model weights remain under their original EleutherAI license; the LoRA adapters and BitFit deltas published here are the only artefacts covered by this repo's license.