| --- |
| library_name: peft |
| license: other |
| tags: |
| - peft |
| - lora |
| - tinker |
| --- |
| |
| # EM code subliminal transfer checkpoints |
|
|
| Version `1.0.0`. This repository contains 34 PEFT LoRA inference |
| checkpoints. |
|
|
| ## Base models |
|
|
| - `Qwen/Qwen3-30B-A3B-Instruct-2507` |
| - `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16` |
|
|
| ## Layout |
|
|
| ```text |
| checkpoints/teacher/ |
| checkpoints/qwen/<condition>/epoch-<1|2>/ |
| checkpoints/nemotron/<condition>/epoch-<1|2>/ |
| ``` |
|
|
| Core epoch checkpoints are steps 1,815 and 3,630. Control checkpoints use the |
| same exposure-matched steps. The dataset-generation teacher is GRPO step 525. |
|
|
| Each checkpoint directory contains `adapter_config.json`, |
| `adapter_model.safetensors`, and `metadata.json`. Exact provenance and SHA-256 |
| values are in `manifest.json`. |
|
|
| ## Loading |
|
|
| ```python |
| from peft import PeftModel |
| from transformers import AutoModelForCausalLM |
| |
| repo = "rustem17/em-code-subliminal-transfer-checkpoints" |
| subfolder = "checkpoints/qwen/insecure_all_comments_removed/epoch-2" |
| base = "Qwen/Qwen3-30B-A3B-Instruct-2507" |
| |
| model = AutoModelForCausalLM.from_pretrained(base) |
| model = PeftModel.from_pretrained(model, repo, subfolder=subfolder) |
| ``` |
|
|
