From 4b9858868fede8cf6d9dff82f792f792e5fb4e72 Mon Sep 17 00:00:00 2001 From: Claude Date: Tue, 9 Jun 2026 03:31:05 +0000 Subject: [PATCH 1/4] Update Zenodo dataset reference to record 20603331 --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 5d91b25..99c887b 100644 --- a/README.md +++ b/README.md @@ -128,12 +128,12 @@ Our setup uses `torch==2.8.0` on the following hardware. ### 3. Download the data The curated DRIP training and evaluation data is archived on Zenodo -(DOI: [10.5281/zenodo.20325769](https://doi.org/10.5281/zenodo.20325769)). +(DOI: [10.5281/zenodo.20603331](https://doi.org/10.5281/zenodo.20603331)). Download and extract it into the repository root: ```bash -wget -O datasets.zip "https://zenodo.org/records/20325769/files/datasets.zip?download=1" +wget -O datasets.zip "https://zenodo.org/records/20603331/files/datasets.zip?download=1" unzip datasets.zip mv datasets1/ datasets/ # rename it to datasets/ ``` From 7d02cfb23c1dd715cb8793423ee0614fa4570346 Mon Sep 17 00:00:00 2001 From: Claude Date: Mon, 15 Jun 2026 02:02:47 +0000 Subject: [PATCH 2/4] Document pretrained DRIP checkpoints: download + merge instructions --- README.md | 60 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 60 insertions(+) diff --git a/README.md b/README.md index 99c887b..3d69696 100644 --- a/README.md +++ b/README.md @@ -25,6 +25,7 @@ DRIP introduces two architectural modifications: - [1. Download base model checkpoints](#1-download-base-model-checkpoints) - [2. Create the environment](#2-create-the-environment) - [3. Download the data](#3-download-the-data) + - [4. (Optional) Download pretrained checkpoints](#4-optional-download-pretrained-checkpoints) - [Training](#training) - [Evaluation](#evaluation) - [SEP score](#sep-score) @@ -144,6 +145,65 @@ curated files used by the training and evaluation scripts. To regenerate the DRIP training data from scratch instead, see [`data_generation/README.md`](./data_generation/README.md). +### 4. (Optional) Download pretrained checkpoints + +If you would rather skip training, we release the DRIP adapters on the Hugging +Face Hub. They are published as **LoRA adapters**, so after downloading you must +**merge** each one into its base model before evaluation (this produces the full +checkpoint the eval scripts load). + +| Checkpoint | Base model | Template | Tool calls | Hugging Face | +|---|---|---|---|---| +| Alpaca + InjecAgent | `meta-llama/Llama-3.1-8B-Instruct` | 4-role (`TextTextText-4roles`) | ✅ supported | [`Kelsey98/Llama-3.1-8B-Instruct-TextTextText-4roles-toolcall-drip`](https://huggingface.co/Kelsey98/Llama-3.1-8B-Instruct-TextTextText-4roles-toolcall-drip) | +| SEP | `meta-llama/Meta-Llama-3-8B-Instruct` | 3-role (`TextTextText`) | — | [`Kelsey98/Meta-Llama-3-8B-Instruct-TextTextText-drip`](https://huggingface.co/Kelsey98/Meta-Llama-3-8B-Instruct-TextTextText-drip) | +| Alpaca | `mistralai/Mistral-7B-Instruct-v0.3` | 3-role (`TextTextTextMistral`) | — | [`Kelsey98/Mistral-7B-Instruct-v0.3-TextTextTextMistral-drip`](https://huggingface.co/Kelsey98/Mistral-7B-Instruct-v0.3-TextTextTextMistral-drip) | + +**Download** the adapter you want (into a local directory), then **merge** it. The +`--base_model_path` and `--customized_model_class` must match the base model and +model family of that checkpoint. + +**Llama-3.1-8B-Instruct · Alpaca + InjecAgent · 4-role / tool-calling** + +```bash +huggingface-cli download Kelsey98/Llama-3.1-8B-Instruct-TextTextText-4roles-toolcall-drip \ + --local-dir meta-llama/Llama-3.1-8B-Instruct-TextTextText-4roles-toolcall-drip + +CUDA_VISIBLE_DEVICES=0 python -m training.merge_lora \ + --adapter_path meta-llama/Llama-3.1-8B-Instruct-TextTextText-4roles-toolcall-drip/ \ + --output_path meta-llama/Llama-3.1-8B-Instruct-TextTextText-4roles-toolcall-drip-merged/ \ + --base_model_path meta-llama/Llama-3.1-8B-Instruct \ + --customized_model_class LlamaForCausalLMDRIP +``` + +**Meta-Llama-3-8B-Instruct · SEP · 3-role** + +```bash +huggingface-cli download Kelsey98/Meta-Llama-3-8B-Instruct-TextTextText-drip \ + --local-dir meta-llama/Meta-Llama-3-8B-Instruct-TextTextText-drip + +CUDA_VISIBLE_DEVICES=0 python -m training.merge_lora \ + --adapter_path meta-llama/Meta-Llama-3-8B-Instruct-TextTextText-drip/ \ + --output_path meta-llama/Meta-Llama-3-8B-Instruct-TextTextText-drip-merged/ \ + --base_model_path meta-llama/Meta-Llama-3-8B-Instruct \ + --customized_model_class LlamaForCausalLMDRIP +``` + +**Mistral-7B-Instruct-v0.3 · Alpaca · 3-role** + +```bash +huggingface-cli download Kelsey98/Mistral-7B-Instruct-v0.3-TextTextTextMistral-drip \ + --local-dir mistralai/Mistral-7B-Instruct-v0.3-TextTextTextMistral-drip + +CUDA_VISIBLE_DEVICES=0 python -m training.merge_lora \ + --adapter_path mistralai/Mistral-7B-Instruct-v0.3-TextTextTextMistral-drip/ \ + --output_path mistralai/Mistral-7B-Instruct-v0.3-TextTextTextMistral-drip-merged/ \ + --base_model_path mistralai/Mistral-7B-Instruct-v0.3 \ + --customized_model_class MistralForCausalLMDRIP +``` + +Pass the **merged** path (`...-merged/`) as the model path in the +[evaluation](#evaluation) scripts. + --- ## Training From 0d69f00880bca0e882c31d79a8d1ae193175d128 Mon Sep 17 00:00:00 2001 From: Claude Date: Mon, 15 Jun 2026 02:06:30 +0000 Subject: [PATCH 3/4] Training table: add dataset column and Alpaca script variants --- README.md | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-) diff --git a/README.md b/README.md index 3d69696..c0532cb 100644 --- a/README.md +++ b/README.md @@ -208,12 +208,17 @@ Pass the **merged** path (`...-merged/`) as the model path in the ## Training -Pick the script that matches your base model: - -| Base model | Command | -|---|---| -| Meta-Llama-3-8B-Instruct | `bash ./scripts/llama8b/sep/drip_sep.sh` | -| Mistral-7B-Instruct-v0.3 | `bash ./scripts/mistral7b/sep/drip_sep.sh` | +Pick the script that matches your **base model** and the **dataset** you want to +train on. The scripts are grouped into per-dataset folders (`sep/`, `alpaca/`) — +to train on SEP run the `sep/` script, to train on Alpaca go to the `alpaca/` +folder and run the matching one there: + +| Base model | Dataset | Command | +|---|---|---| +| Meta-Llama-3-8B-Instruct | SEP | `bash ./scripts/llama8b/sep/drip_sep.sh` | +| Meta-Llama-3-8B-Instruct | Alpaca (3-role) | `bash ./scripts/llama8b/alpaca/drip_alpaca.sh` | +| Meta-Llama-3-8B-Instruct | Alpaca + InjecAgent (4-role / tool-calling) | `bash ./scripts/llama8b/alpaca/drip_alpaca_4roles.sh` | +| Mistral-7B-Instruct-v0.3 | SEP | `bash ./scripts/mistral7b/sep/drip_sep.sh` | Training merges the LoRA adapter into the base weights and saves a **full checkpoint** that evaluation can load directly. For checkpoints saved as adapters From 13f63e87d8bede71df3a669de69b6a911968332c Mon Sep 17 00:00:00 2001 From: Claude Date: Mon, 15 Jun 2026 02:09:37 +0000 Subject: [PATCH 4/4] Fix base model for Alpaca scripts: Llama-3.1-8B-Instruct --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index c0532cb..c2b6e22 100644 --- a/README.md +++ b/README.md @@ -216,8 +216,8 @@ folder and run the matching one there: | Base model | Dataset | Command | |---|---|---| | Meta-Llama-3-8B-Instruct | SEP | `bash ./scripts/llama8b/sep/drip_sep.sh` | -| Meta-Llama-3-8B-Instruct | Alpaca (3-role) | `bash ./scripts/llama8b/alpaca/drip_alpaca.sh` | -| Meta-Llama-3-8B-Instruct | Alpaca + InjecAgent (4-role / tool-calling) | `bash ./scripts/llama8b/alpaca/drip_alpaca_4roles.sh` | +| Llama-3.1-8B-Instruct | Alpaca (3-role) | `bash ./scripts/llama8b/alpaca/drip_alpaca.sh` | +| Llama-3.1-8B-Instruct | Alpaca + InjecAgent (4-role / tool-calling) | `bash ./scripts/llama8b/alpaca/drip_alpaca_4roles.sh` | | Mistral-7B-Instruct-v0.3 | SEP | `bash ./scripts/mistral7b/sep/drip_sep.sh` | Training merges the LoRA adapter into the base weights and saves a **full