Huggggooo/ProtoCycle-Data

---
license: apache-2.0
task_categories:
- text-generation
tags:
- protein-design
- agentic
- tool-use
- reinforcement-learning
language:
- en
size_categories:
- 1K<n<10K
---

# ProtoCycle-Data

Training data for **ProtoCycle** — an agentic protein design model that performs multi-step, tool-augmented sequence design via reinforcement learning.

See the [ProtoCycle](https://github.com/huggggoooooo/ProtoCycle) repository for code, training recipes, and evaluation.

## Dataset Structure

### SFT Data (`sft/desc2seq_agentic_sft_2000.parquet`)

**2,000 multi-turn agentic trajectories** for cold-start supervised fine-tuning.

| Column | Type | Description |
|--------|------|-------------|
| `messages` | list[dict] | Multi-turn conversation with `user`, `assistant`, and `tool` roles. The assistant uses `<think>`, `<plan>`, `<tool_call>`, and `<answer>` tags. |
| `tools` | list[dict] | Tool schemas (10 biology tools: scaffold retrieval, constraint building, ESM inpainting, ProTrek scoring). |

Each trajectory demonstrates the full agent protocol: the model receives a natural-language protein design requirement, reasons step-by-step, invokes biology tools across three stages (scaffold retrieval → constraint injection → refinement & scoring), and outputs a final amino-acid sequence.

### RL Data (`rl/desc2seq_agent_grpo_10000.parquet`)

**10,000 prompts** for GRPO-TCR (Group Relative Policy Optimization with Tool-Call Reward) training.

| Column | Type | Description |
|--------|------|-------------|
| `data_source` | str | Data source identifier (`ProteinDesignEval`) |
| `prompt` | list[dict] | System + user prompt messages for the agent |
| `ability` | str | Task type (`PROTEIN`) |
| `reward_model` | dict | Ground truth and metadata for reward computation |
| `agent_name` | str | Agent type (`tool_agent`) |
| `requirement` | str | Natural-language protein design requirement |
| `requirement_id` | int | Unique requirement identifier |

## Usage

```python
from datasets import load_dataset

# Load SFT data
sft_data = load_dataset("Huggggooo/ProtoCycle-Data", data_dir="sft", split="train")

# Load RL data
rl_data = load_dataset("Huggggooo/ProtoCycle-Data", data_dir="rl", split="train")
```

Or directly with pandas:

```python
import pandas as pd

sft = pd.read_parquet("hf://datasets/Huggggooo/ProtoCycle-Data/sft/desc2seq_agentic_sft_2000.parquet")
rl = pd.read_parquet("hf://datasets/Huggggooo/ProtoCycle-Data/rl/desc2seq_agent_grpo_10000.parquet")
```

## Related Resources

| Resource | Link |
|----------|------|
| ProtoCycle-7B (RL checkpoint) | [Huggggooo/ProtoCycle-7B](https://huggingface.co/Huggggooo/ProtoCycle-7B) |
| ProtoCycle-7B-SFT (SFT checkpoint) | [Huggggooo/ProtoCycle-7B-SFT](https://huggingface.co/Huggggooo/ProtoCycle-7B-SFT) |
| Code & Recipes | [ProtoCycle GitHub](https://github.com/huggggoooooo/ProtoCycle) |

## License

Apache-2.0, consistent with the upstream [VeRL](https://github.com/volcengine/verl) / [Open-AgentRL](https://github.com/Gen-Verse/Open-AgentRL) projects.