History-to-Adapter

History-to-Adapter (H2A) Checkpoints

This repository hosts the six main H2A checkpoints for personalized long-form generation on LongLaMP. H2A learns a shared history-to-weight mapping across users: it compresses a variable-length user history into fixed memory tokens and generates location-specific LoRA weights for Qwen/Qwen2.5-3B-Instruct.

The training and evaluation code is available at xiaolonghan2000/History-to-Adapter.

Checkpoints

Each checkpoint contains the trained H2A History Encoder, Adapter Decoder, projection heads, and checkpoint metadata. The frozen base-model and embedding-model weights are not included.

Download

Download a single checkpoint with the Hugging Face CLI:

hf download Xiaolong-Han/History-to-Adapter \
  checkpoints/topic_writing_user_best_rougeL.pt \
  --local-dir .

Download all six checkpoints:

hf download Xiaolong-Han/History-to-Adapter \
  --include "checkpoints/*.pt" \
  --local-dir .

Evaluation

Clone the H2A source repository, precompute the corresponding LongLaMP history embeddings, and pass the downloaded file to evaluate.py. For example:

python precompute_embeddings.py \
  --config configs/topic_writing_user.yaml \
  --splits test \
  --device cuda

python evaluate.py \
  --config configs/topic_writing_user.yaml \
  --checkpoint checkpoints/topic_writing_user_best_rougeL.pt \
  --split test \
  --device cuda

Full installation, configuration, and reproduction instructions are provided in the GitHub README.

Model Details

• Base model: Qwen/Qwen2.5-3B-Instruct
• History embedding model: Qwen/Qwen3-Embedding-4B
• Memory tokens: 32
• Generated LoRA rank: 8
• Target modules: q_proj, v_proj
• Evaluation gate: 0.3

These checkpoints are tied to the H2A architecture and matching task configurations in the public source repository; they are not standalone Transformers or PEFT checkpoints.