metadata
base_model: unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit
library_name: peft
pipeline_tag: text-generation
tags:
- lora
- grpo
- trl
- unsloth
- quantum-error-correction
license: mit
QuantumScribe (GRPO LoRA)
LoRA adapter fine-tuned with GRPO for logical quantum error correction, on top of base unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit.
Adapter
- LoRA
r=16,lora_alpha=32,lora_dropout=0.1 - Target:
q_proj,k_proj,v_proj,o_proj(PEFT 0.18.1)
Training
- W&B: ronitraj/QuantumScribe-GRPO (e.g. run
4p7eurnc) - ~1500 GRPO steps; SFT warm-up as in the project
scripts/train_grpo.py
Eval (from project data/eval_grpo.json)
- Logical correction rate high (~0.96 on the recorded run)
- pymatching_beat reported at 0 on the evaluated split — align narrative and metrics (continuous vs threshold) with your harness and README
Load
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_id = "unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit"
adapter_id = "ronitraj/quantumscribe"
tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(
base_id, device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(model, adapter_id)