README.md

---
license: mit
language:
- en
pipeline_tag: text-generation
tags:
- chemistry
- covalent-organic-frameworks
- reactive-oxygen-species
- redox
- qlora
- sft
- scientific-qa
datasets:
- Willlzh/COFOS_data
library_name: transformers
---

# COFOS

COFOS is a domain-adapted language model for question answering about covalent
organic frameworks (COFs), reactive oxygen species (ROS), oxygen-derived
products, and photocatalytic or redox reaction mechanisms.

The model is intended to answer chemistry questions in a natural QA style while
preserving important distinctions such as:

- dominant ROS or oxygen-derived products versus secondary ROS or intermediates
- H2O2 as an oxygen-derived product rather than a radical ROS
- condition-dependent behavior under light, oxygen, water, sacrificial reagents,
  PMS, or related reaction environments
- uncertainty when the available information is insufficient

## Model Details

- **Model name:** COFOS
- **Repository:** `Willlzh/COFOS`
- **Model type:** causal language model
- **Architecture family:** Qwen-style decoder-only language model
- **Training method:** QLoRA SFT followed by adapter merge
- **Training data:** [`Willlzh/COFOS_data`](https://huggingface.co/datasets/Willlzh/COFOS_data)
- **Primary language:** English
- **License:** MIT

The uploaded checkpoint is a merged Transformers model directory. It can be
loaded directly with `AutoModelForCausalLM.from_pretrained()` without separately
loading a LoRA adapter.

## Intended Use

COFOS is designed for research-oriented QA and drafting assistance around:

- COF photocatalysis
- ROS generation and assignment
- oxygen reduction and H2O2 photoproduction
- PMS-assisted oxidation mechanisms
- evidence-aware explanations of dominant versus secondary species
- chemistry and redox QA in an educational or literature-review setting

It is not a substitute for experimental validation, safety review, or expert
chemical judgment.

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Willlzh/COFOS"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

question = "What ROS does TT-T-COF generate under visible-light photocatalysis?"
messages = [
    {
        "role": "system",
        "content": (
            "You are COFOS, a natural QA assistant for covalent organic "
            "frameworks and reactive oxygen species. Answer directly and do "
            "not call H2O2 a radical ROS."
        ),
    },
    {"role": "user", "content": question},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Example Questions

```text
What ROS does TT-T-COF generate under visible-light photocatalysis?
```

```text
In a Fe3O4@TpMa/PMS photocatalytic system, which ROS are dominant for phenol degradation?
```

```text
Why should H2O2 not be described as a radical ROS?
```

```text
What should be considered before assigning O2·- or ·OH as the dominant ROS?
```

## Training Data

The model was trained with a mixture of COFOS-specific and chemistry-oriented
SFT data:

- `cofos_teacher_distill.jsonl`: teacher-distilled COF/ROS QA data
- `cofos_rag_sft.jsonl`: RAG-style samples with question, KG facts, retrieved
  evidence, and answer
- `chem_redox_sft.jsonl`: English chemistry/redox QA samples
- `style_correction_sft.jsonl`: answer-style correction samples focused on
  natural QA behavior and avoiding awkward evidence boilerplate

The dataset is available at
[`Willlzh/COFOS_data`](https://huggingface.co/datasets/Willlzh/COFOS_data).

## Limitations

- The model may still hallucinate details for materials or conditions that are
  not represented in its training data.
- The model is not guaranteed to cite sources correctly unless used with an
  external retrieval system.
- Mechanistic explanations should be treated as research assistance rather than
  final experimental conclusions.
- The model should not be used for high-stakes chemical safety, medical, legal,
  or regulatory decisions.

## Recommended Deployment Pattern

For the best COFOS experience, use the model as part of a local persistent
assistant or a retrieval-augmented workflow:

```text
user question
-> optional KG/BM25 retrieval
-> COFOS model
-> natural QA answer
```

The merged model can also be used directly for lightweight QA without retrieval,
but RAG is recommended for record-specific material questions.

## Citation

If you use COFOS in a project, please cite or link this model repository and the
associated dataset:

- Model: `Willlzh/COFOS`
- Dataset: `Willlzh/COFOS_data`