PolyVL-logic-lora
an adapter for logic tasks and deduction
1 min readupdated
This PEFT LoRA (a lightweight add-on for a base model) adapter improves formal deduction, propositional logic, and puzzle-solving. Built on top of the OpenGVLab/InternVL3-2B language model, it was trained on a dedicated dataset combining Dolly-15k, ProofWriter, FOLIO, and synthetic logic problems. The model improves accuracy on multi-step logical inference and symbolic constraints, making it a solid fit for researchers and developers working on logical reasoning in compact models.
Key features
- Built on InternVL3-2B using the peft library.
- Trained on datasets for formal and natural deduction.
- Configured with LoRA rank r=64, alpha=128, and a 0.05 dropout.
- Adapter weights are released under the Apache-2.0 license.
What you can do
- Solve complex logic problems and puzzles.
- Handle natural language inference and deductive analysis tasks.
- Integrate into text processing pipelines to verify logical consistency.
Example
import torch
from transformers import AutoModel, AutoTokenizer
from peft import PeftModel
BASE_MODEL = "OpenGVLab/InternVL3-2B"
BASE_REVISION = "899155015275a9b7338c7f4677e19c784e0e5a21"
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, revision=BASE_REVISION, trust_remote_code=True)
base_model = AutoModel.from_pretrained(
BASE_MODEL,
revision=BASE_REVISION,
torch_dtype=torch.bfloat16,
trust_remote_code=True
).eval()
model = PeftModel.from_pretrained(base_model, "Pushkarrrr15151/PolyVL-logic-lora").eval()Read more
DetailsBenchmarks: LogicBench, LogicNLI, Multi-LogiEval
Base model: OpenGVLab/InternVL3-2B
Weights license: Apache-2.0










Comments