Modelpublic

ho22joshua/hep-posttraining

sourceHugging Faceotherupdated 3d agoView on Hugging Face

0likes

Model Card

HEP post-training adapters

This repository contains standalone PEFT/LoRA adapters for HEP analysis tasks. It deliberately excludes resumable FSDP checkpoints, optimizer states, and training data. Each adapter directory includes adapter_model.safetensors, adapter_config.json, tokenizer assets, and the Qwen chat template.

Adapter catalog

Task	Adapter directory	Base model	LoRA	Epochs	Learning rate
ROOT Dataset	`ROOT/qwen3.5-0.8b-lora-r32-e5-lr1e-5`	`Qwen/Qwen3.5-0.8B`	32/32	5	1e-5
ROOT Dataset	`ROOT/qwen3.5-4b-lora-r32-e1-lr1e-4`	`Qwen/Qwen3.5-4B`	32/32	1	1e-4
ROOT Dataset	`ROOT/qwen3.5-9b-lora-r32-e10-lr1e-5`	`Qwen/Qwen3.5-9B`	32/32	10	1e-5
ROOT Dataset	`ROOT/qwen3.5-9b-lora-r32-e1-lr1e-4`	`Qwen/Qwen3.5-9B`	32/32	1	1e-4
ROOT + TRExFitter Dataset	`ROOT-plus-TRExFitter/qwen3.5-0.8b-lora-r16-e3-lr1e-5`	`Qwen/Qwen3.5-0.8B`	16/16	3	1e-5
ROOT + TRExFitter Dataset	`ROOT-plus-TRExFitter/qwen3.5-0.8b-lora-r32-e3-lr1e-4`	`Qwen/Qwen3.5-0.8B`	32/32	3	1e-4

There is currently no finished, standalone TRExFitter-only adapter. Existing TRExFitter-only experiments are incomplete RL checkpoints and are intentionally not published.

Load an adapter

python

import torch
from peft import PeftModel
from transformers import AutoModelForMultimodalLM, AutoProcessor

base_id = "Qwen/Qwen3.5-4B"
adapter_dir = "ROOT/qwen3.5-4b-lora-r32-e1-lr1e-4"

processor = AutoProcessor.from_pretrained(adapter_dir)
base = AutoModelForMultimodalLM.from_pretrained(
    base_id, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, adapter_dir).eval()

Clone with Git LFS so adapter weights are downloaded:

bash

git lfs install
git clone https://huggingface.co/ho22joshua/hep-posttraining

The adapter configuration records its required base model. Use the included chat_template.jinja and tokenizer/processor files when constructing prompts.