selfplay-llama-p5en-130d5b

SelfPlay A1 checkpoints

Run: p5en-e6c22491-l-130d5b. Topic-label model: local Llama.

These are LoRA adapters for the Llama-3.1-8B base model, saved during the first attacker training stage (PSRO/A1). GPT, where named, supplied topic labels; these files do not contain GPT model weights.

Saved steps in this repository: 160, 170, 180.

Training status: A1 terminated by the original COLLAPSE guard at step 180 after sustained output-format violations. The full PSRO/Uniform/Naive experiment did not complete, and independent benchmarks were not run. Selection follows the largest step count, not a best-quality checkpoint criterion.

Load a selected checkpoint

Install compatible torch, transformers, peft, and accelerate packages. The base model must also be downloaded; it is not included in this repository.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated"
base_revision = "368c8ed94ce4c986e7b9ca5c159651ef753908ce"
repo_id = "zhihz0535/selfplay-llama-p5en-130d5b"
checkpoint = "checkpoints/step-180"

tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)
base = AutoModelForCausalLM.from_pretrained(
    base_id, revision=base_revision, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(base, repo_id, subfolder=checkpoint)

The checkpoints in this repository are publicly accessible.

run-info.json records the settings and export changes. manifest.json contains file sizes and SHA-256 checksums. Original adapter configurations are retained under provenance/. Adapter weights were not changed.