README_hf.md

metadata

license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
library_name: peft
tags:
  - lora
  - peft
  - qwen2.5
  - multi-agent
  - cooperative
  - overcooked
language:
  - en

DriCo: Planner / Actor / Coordinator LoRA Adapters

Four LoRA adapters fine-tuned on Qwen/Qwen2.5-7B-Instruct for a multi-agent cooperative-cooking benchmark.

A single base model is loaded once at inference; the three roles are switched by activating the matching adapter.

Folder	Role	Training
planner/	High-level subgoal generation in natural language (e.g. "Pick up carrot and put it in pot0.")	SFT
actor/	Low-level action chunking — turns a subgoal into up to 3 ml_actions	SFT
coordinator/	Shared-context generation + PASS/REJECT critique of planner subgoals	SFT
coordinator_dpo/	Coordinator further refined with DPO	SFT → DPO

---

Quick start

End-to-end: clone the project repo, drop these adapters into the right folder, run eval.

# 1) Clone the project repo
git clone https://github.com/anonymous-projectpage/DriCo.git
cd DriCo

# 2) Install deps (in a fresh env)
pip install -U huggingface_hub
pip install -r requirements.txt          # if the repo has one

# 3) Download these adapters into drico/out/
hf download anonymous-24421/DriCo-adapters \
    --local-dir drico/out

# 4) (Optional) download the training/DPO data into drico/dataset/
hf download anonymous-24421/DriCo \
    --repo-type dataset \
    --local-dir drico/dataset

# 5) Run evaluation
cd drico
CUDA_VISIBLE_DEVICES=0 python eval_proposed.py \
    --base_model           Qwen/Qwen2.5-7B-Instruct \
    --adapter_root         out/ \
    --layout               circuit_env \
    --recipe_split         test \
    --statistics_save_dir  data/eval_test

After step 3, the directory tree will look like:

DriCo/
└── drico/
    └── out/
        ├── README.md          ← this file
        ├── planner/
        │   ├── adapter_config.json
        │   ├── adapter_model.safetensors
        │   └── ...
        ├── actor/
        │   └── ...
        ├── coordinator/
        │   └── ...
        └── coordinator_dpo/
            └── ...

That's exactly what eval_proposed.py --adapter_root out/ expects.

---

Selective download

Only the actor adapter:

hf download anonymous-24421/DriCo-adapters \
    --include "actor/*" \
    --local-dir drico/out

Only the DPO-refined coordinator:

hf download anonymous-24421/DriCo-adapters \
    --include "coordinator_dpo/*" \
    --local-dir drico/out

---

Manual loading in Python

If you want to use the adapters outside the eval script:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen2.5-7B-Instruct"
tok  = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base, torch_dtype="bfloat16", device_map="auto",
)

# Register the three roles as named adapters on one base model
model = PeftModel.from_pretrained(
    model, "anonymous-24421/DriCo-adapters",
    subfolder="planner", adapter_name="planner",
)
model.load_adapter("anonymous-24421/DriCo-adapters",
                   subfolder="actor", adapter_name="actor")
model.load_adapter("anonymous-24421/DriCo-adapters",
                   subfolder="coordinator_dpo", adapter_name="coordinator")

# Switch role per call
model.set_adapter("planner")
out = model.generate(
    **tok("Plan the next subgoal: ...", return_tensors="pt").to(model.device),
    max_new_tokens=50,
)
print(tok.decode(out[0], skip_special_tokens=True))

model.set_adapter("actor")
# ... generate the next action chunk ...

Swap subfolder="coordinator_dpo" for subfolder="coordinator" to use the SFT-only coordinator instead of the DPO-refined one.

---

Related

• Project repo: https://github.com/anonymous-projectpage/DriCo
• Training data: https://huggingface.co/datasets/anonymous-24421/DriCo
• Base model: https://huggingface.co/Qwen/Qwen2.5-7B-Instruct

---

Intended use

Research on multi-agent cooperative task planning, role-conditioned generation, and critique–refine pipelines. Not intended for general chat/assistant use.

Limitations

• Trained for a specific cooperative-cooking environment with a fixed action vocabulary (pickup, put_obj_in_utensil, cook, bake, stir, ...).
• Outputs outside that domain are not guaranteed to be coherent.
• The coordinator's critique is a heuristic signal, not a guarantee of task correctness.

Citation

@misc{drico2026,
  title  = {DriCo: Planner / Actor / Coordinator for Cooperative Cooking},
  author = {Anonymous},
  year   = {2026},
  note   = {Under review}
}