parallel-ttt-cot-qwen3-1b-v1

This repository contains a standalone PEFT LoRA adapter exported from the Lightning checkpoint epoch=0-step=4663.ckpt. The adapter is not merged into the base model, so it can be loaded later and continued from as a trainable LoRA adapter.

Important: although the repo name says qwen3-1b, the checkpoint itself was trained against the base model Qwen/Qwen3-1.7B and the adapter metadata points to that base model.

What is in this repo

• PEFT adapter files: adapter_model.safetensors, adapter_config.json
• Base-model text assets: tokenizer.json, tokenizer_config.json, chat_template.jinja, generation_config.json, config.json
• Extra non-PEFT trainables from the original checkpoint: non_peft_trainables.safetensors
• Lightning-loadable checkpoint for module_class.load_from_checkpoint(..., config=args, ...): lightning_checkpoint.ckpt
• Metadata: export_metadata.json, non_peft_trainables_metadata.json, lightning_checkpoint_metadata.json

Full trainable state summary

The original training state exported here includes:

• 56 LoRA tensors saved in adapter_model.safetensors
• 4 weighting-model tensors in non_peft_trainables.safetensors:
  • weight_model.fc1.weight
  • weight_model.fc1.bias
  • weight_model.fc2.weight
  • weight_model.fc2.bias
• 1 learned inner-loop LR tensor in non_peft_trainables.safetensors:
  • lrs.lr

Load adapter for continued LoRA finetuning

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen3-1.7B"
adapter_id = "zechen-nlp/parallel-ttt-cot-qwen3-1b-v1"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id, is_trainable=True)

Load the extra non-PEFT trainables

from safetensors.torch import load_file

extra = load_file("non_peft_trainables.safetensors")
weight_model_state = {
    k.split('weight_model.', 1)[1]: v
    for k, v in extra.items()
    if k.startswith('weight_model.')
}
learned_inner_lrs = extra['lrs.lr']

Load the Lightning checkpoint

This checkpoint is intended for the same loading path used in perk/runner.py:

model = module_class.load_from_checkpoint(
    checkpoint_path,
    config=args,
    map_location="cuda:0",
)

If you want to download it from the Hub first:

from huggingface_hub import hf_hub_download
from omegaconf import OmegaConf
from perk.modules import MetaLearningModule

repo_id = "zechen-nlp/parallel-ttt-cot-qwen3-1b-v1"
checkpoint_path = hf_hub_download(repo_id=repo_id, filename="lightning_checkpoint.ckpt")

# Supply a compatible config explicitly. The checkpoint intentionally omits
# hyper_parameters and other training/runtime state.
args = OmegaConf.create({
    # fill in the same config fields your module expects
})

model = MetaLearningModule.load_from_checkpoint(
    checkpoint_path,
    config=args,
    map_location="cpu",
)

lightning_checkpoint.ckpt only keeps the top-level keys needed for that load path:

• state_dict
• pytorch-lightning_version

So unlike a full Lightning training checkpoint, it does not include optimizer state, scheduler state, callbacks, loops, or embedded hyperparameters.