YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
260326_local_gpu_training — MO3 dual-LoRA collusion organism: trained LoRAs and merges
Llama-3.3-70B-Instruct LoRA adapters for the MO3 dual-LoRA collusion organism (policy LoRA + monitor LoRA, trained
with Tinker), plus five ways of merging them into a single adapter. Backed up 2026-10-02 from RunPod pod
574248e2fc5a (213.173.108.41); work dated 2026-03-26 → 2026-03-31.
Code, merge scripts (merge_*.py), eval scripts and RESULTS.md: GitHub jprivera44/collusion_project_v0 @
mo3-local-gpu-merge-experiments, folder experiments/260326_local_gpu_training/. Claude chats for this work:
jprivera/private_backup → 260330_mo3_merge_experiments/ (private).
| Path | What | Original pod dir (experiments/260326_local_gpu_training/output/) |
|---|---|---|
output/mo3_policy_lora_4b87213b/ |
Policy LoRA (Tinker sampler weights) | 4b87213b-3c56-55f7-b07b-a5dd15f87ad3:train:0_sampler_weights_mo3_policy_lora/ |
output/mo3_monitor_lora_ef0720f6/ |
Monitor LoRA (Tinker sampler weights) | ef0720f6-cf75-5a2e-b8d8-3cc5c7afdf7e:train:0_sampler_weights_mo3_monitor_lora/ |
output/mo3_merged_lora/ |
Weight-sum merge (merge_lora.py) |
mo3_merged_lora/ |
output/mo3_cat_merge/ |
Concatenated-rank merge (merge_cat.py) |
mo3_cat_merge/ |
output/mo3_scaled_merge/ |
Scaled merge (merge_scaled.py) |
mo3_scaled_merge/ |
output/mo3_ties_merge/ |
TIES merge (merge_ties.py) |
mo3_ties_merge/ |
output/mo3_correct_merge/ |
SVD-based merge (merge_correct.py, svd_stats.json) |
mo3_correct_merge/ |
Directory names were shortened because the originals contain :. MANIFEST.tsv has sha256 for every file.