PoRT IPC T5

Recovered AST-generation checkpoint used by the In-Context Prompt Cleaning (IPC) stage of PoRT: Robust LLM Unlearning via Post Judgment and Multi-Round Thinking.

Code and pipeline: https://github.com/ChnIRuI/PoRT_LLM_Unlearning

Download and load

Install the project dependencies first, then download the complete repository:

from huggingface_hub import snapshot_download
from transformers import T5TokenizerFast, T5ForConditionalGeneration

model_dir = snapshot_download("ChnIRuI111/PoRT-IPC-T5")
tokenizer = T5TokenizerFast.from_pretrained(model_dir)
model = T5ForConditionalGeneration.from_pretrained(model_dir)

Set the PoRT pipeline argument --t5_model_path to model_dir. Download all configuration and tokenizer files together with model.safetensors.

For a private repository, the downloading account must be authorized and logged in to Hugging Face.

Provenance and scope

  • Archived TOFU launch scripts repeatedly point to the recovered t5_ast_finetuned root directory.
  • The retained training source initializes from a local flan-t5-large model and trains query-to-AST generation.
  • This release contains the root inference export, not the nested checkpoint-11100 training directory.
  • The tensor header identifies a T5 encoder-decoder with 24 encoder layers, 24 decoder layers, and float32 weights.
  • The model assists the IPC pipeline by generating AST representations. It is not the response-generation LLM.
  • Recovery checks covered source/script references, checkpoint structure, file sizes, and SHA-256 hashes. Full inference and paper-table reproduction were not rerun for this release.

SHA256SUMS records the exact uploaded payload files.

Upstream

Base model: https://huggingface.co/google/flan-t5-large

The upstream T5 code license is included in LICENSE.

Downloads last month
17
Safetensors
Model size
0.8B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ChnIRuI111/PoRT-IPC-T5

Finetuned
(226)
this model