Instructions to use phenixace/Chem-R-Faithful with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use phenixace/Chem-R-Faithful with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="phenixace/Chem-R-Faithful")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("phenixace/Chem-R-Faithful")
model = AutoModelForCausalLM.from_pretrained("phenixace/Chem-R-Faithful", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use phenixace/Chem-R-Faithful with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "phenixace/Chem-R-Faithful"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "phenixace/Chem-R-Faithful",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/phenixace/Chem-R-Faithful

SGLang

How to use phenixace/Chem-R-Faithful with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "phenixace/Chem-R-Faithful" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "phenixace/Chem-R-Faithful",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "phenixace/Chem-R-Faithful" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "phenixace/Chem-R-Faithful",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use phenixace/Chem-R-Faithful with Docker Model Runner:
```
docker model run hf.co/phenixace/Chem-R-Faithful
```

Chem-R-Faithful

Chem-R-8B continued with GRPO under a verification-grounded process reward: the accuracy term is paid only when the reasoning trace is clean, i.e. when every functional-group claim it makes is supported by the input, the predicted molecule, or the reference.

The point is not accuracy. An answer-level reward cannot distinguish a correct answer reached through a faithful trace from the same answer reached through a fabricating one, because the reward does not depend on the trace at all. Gating it on a structural check of the trace makes those two cases score differently.

What it changes

Measured over twelve generative task variants (ChEBI-20 caption↔molecule, USPTO-50k retrosynthesis, nine S²-Bench subtasks), against the Chem-R checkpoint it was trained from. All four rows below are unweighted means over the twelve task variants (each variant has weight 1/12), calculated from the per-task values in sheet Diagnosis_model_task of the repository's Source Data. This uses the same task-averaging convention as the paper's Fig. 1d and its mean ER across tasks; the task sizes do not determine their weights in this table.

Specifically, task performance, mean ER and clean-trace rate average the perf, ER and pct_er0 columns, respectively. Per-claim fabrication rate averages 100 - cp over tasks, where cp is the percentage of verified claims among verified plus fabricated specific functional-group claims within each task; the six generic group names excluded by the evaluation code are excluded from both counts. These are means of per-task claim rates, not a single claim rate pooled across tasks.

	Chem-R	Chem-R-Faithful
per-claim fabrication rate	22.56%	4.35%
mean ER (fabrication score, 0–100)	10.63	2.39
clean-trace rate (ER = 0)	45.45%	87.58%
task performance (0–100)	58.91	61.48

Values here are shown to two decimal places. Rounded directly from the underlying values to the paper's display precision, performance is 59 / 61 (Fig. 1d), and Chem-R-Faithful has mean ER 2.4 and per-claim fabrication rate 4.3%.

Aggregation update: Earlier versions of this card took performance, mean ER and clean-trace rate from R1_stage_ladder, which weights these metrics by the number of responses (16,107 per model: 3,300 for each ChEBI-20 task, 5,007 for retrosynthesis and 500 for each of the nine S²-Bench subtasks). Those response-weighted summaries gave performance 50.09 / 51.45 and Chem-R-Faithful mean ER 1.73. The table above now uses equal task weights to align with the paper; the model checkpoints, evaluation records and per-task results are unchanged. Per-claim fabrication rate was already task-averaged and is unchanged.

Fabrication drops roughly five-fold and the clean-trace rate nearly doubles, with task performance slightly up rather than traded away.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("phenixace/Chem-R-Faithful", torch_dtype="bfloat16")
t = AutoTokenizer.from_pretrained("phenixace/Chem-R-Faithful")

It answers in <think>…</think><answer>…</answer> form. The answer span is what should be parsed; the trace span is what the detector audits.

Training

GRPO, 936 steps on 4×H200 (~26.6 h). Reward 0.1·format + 0.4·accuracy + 0.4·(1 − hallucination) + 0.2·grounded, with the accuracy term gated on ER = 0. Config, launcher, the exact training parquets, and the EasyR1/verl patch the per-task reward dispatch requires are in the code repository.

Code, data, and the detector

https://github.com/phenixace/MolReHallu — the structural claim detector, the evaluation pipeline, the released model responses and per-claim diagnosis records, and the training recipe. The detector and the reward gate can both be exercised on a CPU in under a minute.

Limitations

The verifier decides functional groups, ring systems and molecular classes from the molecular graph. It does not certify a complete chemical argument, and ER is a fabrication rate over the explicit, structurally decidable claims it recovers rather than a recall-complete audit of the reasoning. Fabrication is reduced, not eliminated.

License

CC BY 4.0. Derived from weidawang/Chem-R-8B, which remains subject to its own license.

Downloads last month: 20

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for phenixace/Chem-R-Faithful

Base model

meta-llama/Llama-3.1-8B

Finetuned

meta-llama/Llama-3.1-8B-Instruct

Finetuned

weidawang/Chem-R-8B

Finetuned

(1)

this model