library_name: peft
base_model:
- Qwen/Qwen3-8B
- Qwen/Qwen3-14B
tags:
- peft
- lora
- continual-learning
- alignment
- ai-safety
- model-trajectories
Direction, Not Distance? — Full Continual-Learning Trajectories
This repository contains the complete phase-by-phase LoRA checkpoint trajectories produced for the Direction, Not Distance? experiments.
These are research artifacts, not standalone language models.
Contents
Main experiment
Qwen3-8B:
- seeds 42–48
- benign and alignment-conflicting continual learning
- unconstrained
- global projection
- coordinate mortality
- phases 0–6
Qwen3-14B:
- seeds 42–44
- benign and alignment-conflicting continual learning
- unconstrained
- global projection
- coordinate mortality
- phases 0–6
Targeted follow-up
Qwen3-8B:
- seeds 42–48
- benign and alignment-conflicting continual learning
- norm-matched shrinkage
- phases 0–6
Why release intermediate checkpoints?
The complete trajectories permit independent researchers to:
- reproduce phase-level preference-drift analyses;
- apply alternative behavioral evaluators;
- conduct new mechanistic-interpretability analyses;
- study when coordinate-level interference emerges;
- test alternative probes and alignment metrics;
- independently audit trajectory-level claims.
Related resources
Code, results, figures and experimental manifests:
https://github.com/SubramanyamSahoo/Direction-Not-Distance
Final/preference-tuned LoRA checkpoints:
https://huggingface.co/SahoobhAI/Direction-Not-Distance-LoRA
Integrity
FILE_MANIFEST.txt lists all uploaded files.
SHA256SUMS.txt provides SHA-256 hashes for integrity verification.
Important caveat
The reported likelihood-based preference improvements do not consistently transfer to the independent ArmoRM behavioral evaluator. These checkpoints should not be interpreted as providing a universal behavioral alignment guarantee.