ho22joshua/hep-chat-entrypoint

HEP Posttraining Chat Entrypoint

Run the LoRA adapters published in ho22joshua/hep-posttraining. The repository contains only lightweight entrypoint code: it downloads the selected Qwen 3.5 base model and applies the selected PEFT adapter at runtime.

The registry includes every published adapter:

• ROOT Dataset: Qwen 3.5 0.8B (r32/e5/1e-5), 4B (r32/e1/1e-4), and 9B (r32/e1/1e-4; r32/e10/1e-5).
• ROOT + TRExFitter Dataset: Qwen 3.5 0.8B (r16/e3/1e-5; r32/e3/1e-4).

Install

Use a CUDA GPU for the 4B and 9B models.

git clone https://huggingface.co/ho22joshua/hep-chat-entrypoint
cd hep-chat-entrypoint
python -m pip install -r requirements.txt
python list_models.py

Interactive chat

The default is the one-epoch ROOT Dataset 4B adapter, with thinking disabled:

python chat.py --device cuda

Choose another adapter or compare it to its unadapted base model:

python chat.py --model root-qwen3.5-9b-r32-e1-lr1e-4 --device cuda
python chat.py --model root-qwen3.5-9b-r32-e1-lr1e-4 --mode base --device cuda

Enable the Qwen reasoning mode for an explicit thinking ablation:

python chat.py --model root-qwen3.5-4b-r32-e1-lr1e-4 --thinking --device cuda

During chat, use /thinking on, /thinking off, or /exit.

Batch inference

Put one prompt on each non-empty line of prompts.txt:

python run_prompts.py \
  --model root-qwen3.5-4b-r32-e1-lr1e-4 \
  --prompts prompts.txt \
  --output outputs.jsonl \
  --device cuda

outputs.jsonl begins with run metadata and then one completion record per prompt. Add --thinking for the reasoning condition, or --mode base for the matched base-model condition.

Gradio

python app.py

The UI provides an adapter picker, base-versus-adapter mode, and a thinking toggle. It loads one model at a time.

Direct PEFT usage

Each registry entry specifies the base model and adapter subfolder. The equivalent minimal Python pattern is:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", torch_dtype="auto")
model = PeftModel.from_pretrained(
    base,
    "ho22joshua/hep-posttraining",
    subfolder="ROOT/qwen3.5-4b-lora-r32-e1-lr1e-4",
)