Instructions to use QMJWang/FaithOR-Detector-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use QMJWang/FaithOR-Detector-4B with Transformers:

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

pipe = pipeline("text-generation", model="QMJWang/FaithOR-Detector-4B")
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("QMJWang/FaithOR-Detector-4B")
model = AutoModelForCausalLM.from_pretrained("QMJWang/FaithOR-Detector-4B", 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 QMJWang/FaithOR-Detector-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "QMJWang/FaithOR-Detector-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "QMJWang/FaithOR-Detector-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/QMJWang/FaithOR-Detector-4B

SGLang

How to use QMJWang/FaithOR-Detector-4B 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 "QMJWang/FaithOR-Detector-4B" \
    --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": "QMJWang/FaithOR-Detector-4B",
		"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 "QMJWang/FaithOR-Detector-4B" \
        --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": "QMJWang/FaithOR-Detector-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use QMJWang/FaithOR-Detector-4B with Docker Model Runner:
```
docker model run hf.co/QMJWang/FaithOR-Detector-4B
```

FaithOR-Detector-4B

A Qwen3-4B detector that reads a natural-language optimization problem, a numbered candidate mathematical model and the model's own solver result, and reports which components are semantically wrong, what kind of error each one is, and whether a required constraint is missing. A frozen writer then applies one targeted repair. The detector is a plain causal LM; the diagnosis is generated text, no classification head is added.

Paper: coming soon (arXiv)
Training data: FaithOR-Supervision
Code (training, inference, one-round repair, evaluation): github.com/QMJWang/FaithOR

Input and output

The prompt contains the problem, the solver status and objective of the candidate model, and the model with every component numbered (S sets, P parameters, V variables, O1 objective, C constraints), followed by a fixed task instruction. The response is one short check line and a JSON diagnosis:

Checked 12 components against the problem.
{"errors": [{"id": "V1", "type": "var_to_int"}], "missing": false}

Error types: var_to_int, var_to_bin, var_to_cont, var_bound, con_sense, num_value, con_structure, con_extra, obj_sense, obj_expr. A clean model gets {"errors": [], "missing": false}. Use the prompt builder and parser from the code repository; an unparsable response is treated as an abstention, not as a clean verdict. Thinking mode must be disabled.

Usage

vllm serve QMJWang/FaithOR-Detector-4B --served-model-name faithor --port 8000 --max-model-len 16384

pip install -e .            # in the FaithOR code repository
or-detect --input examples/detector_input.jsonl --output outputs/detection.jsonl --model faithor
or-repair --input examples/problems.jsonl --out outputs/ours \
  --writer-model <writer> --writer-url http://localhost:8001/v1 \
  --coder-model <writer> --coder-url http://localhost:8001/v1 \
  --detector-model faithor --detector-url http://localhost:8000/v1

Decode greedily (temperature 0, max 512 new tokens).

Provenance and training

These weights are the detector evaluated in the paper. They were trained from Qwen/Qwen3-4B on the supervision now published as FaithOR-Supervision, through a sequence of exploratory SFT and GRPO runs; the merged tensor file has 398 tensors, 8,044,982,080 bytes, SHA256 a34bdfb63af9e4a67b7cbf0d43003ebf25bc893f510984bc0d65b618fb7753e0. In the one-round repair harness (Qwen3.6-35B-A3B as writer and coder, seven benchmarks, 1,511 problems, solver-certified optimality with relative tolerance 1e-4) this checkpoint scores 92.55 ± 0.10% over five runs. From those exploratory runs we distilled the fixed three-step pipeline below, released in the code repository; trained from scratch on the same data it scores 1,391 and 1,394 / 1,511 in two runs of the same harness, against 1,398 for these weights in a single run.

The recipe: two supervised stages, then a short GRPO stage. The two SFT stage files are published with the dataset (stages configuration) in the exact row order used here. Both SFT stages are LoRA fine-tuning on the attention and MLP projections (q,k,v,o,gate,up,down), alpha = 2 × rank, dropout 0.05, AdamW without weight decay, 5% warm-up then cosine decay, gradient accumulation 8, one epoch, seed 20261013. The loss is computed on response tokens only, with token weights 10 on the first decision token of the errors list, 3 on component IDs, error types and the missing value, and 1 elsewhere; each example is normalized by its weight sum.

Stage	Data	Rows	LoRA rank	Learning rate	Max length	Steps	Wall time
1. General supervision	`data/sft_stage1.jsonl` (the `optmath` and `resocratic` rows)	16,939	64	1e-4	8192	2,118	3.6 h
2. Semantic contracts	`data/sft_stage2.jsonl` (the 13 simulated families, then the 1,509 `replay` rows)	9,189	16	6e-6	8192	1,149	1.0 h

Stage 2 starts from the merged stage-1 model and its adapter is merged into that model. Stage 1 teaches the protocol and general error recognition from real writer formulations; stage 2 calibrates semantic conventions (implicit integer domains, exact vs. at-least ratios, shared capacities, net-profit objectives, prerequisites) on the simulated contract families while replaying 1,509 real writer formulations so the model does not drift toward the simulator's templates.

Stage 3: GRPO (50 steps). Starting from the merged stage-2 model, full-parameter GRPO with verl 0.9.1 and vLLM 0.30 rollouts on the grpo train split (3,706 prompts over writer-generated formulations): batch 32 prompts, 8 samples per prompt at temperature 1.0, learning rate 1e-6, KL coefficient 0.001 against the stage-2 reference, no entropy bonus, prompt limit 2,400 tokens and response limit 512, FSDP with parameter and optimizer offload on two GPUs. The reward is rule-based (training/grpo_reward.py): −1 for an unparsable response, +1 / −1 for a correct model left alone / wrongly flagged, −0.5 for an erroneous model reported clean, otherwise component-localization F1 plus up to 0.2 for correct error types. Checkpoints were taken every 25 steps; the 50-step checkpoint gave the score above.

Stage	Base	Data	Method	Steps	GPUs	Wall time
3. GRPO	merged stage-2 model	`grpo` train, 3,706 prompts	full-parameter GRPO, lr 1e-6, 32 × 8 rollouts	50	2	2.9 h

Reproduce with the code repository (first restore the upstream problem text in the dataset with scripts/rehydrate.py, see the dataset card):

python training/train.py --config configs/training.json --stage sft1 --base Qwen/Qwen3-4B \
  --data-root ../FaithOR-Supervision --out checkpoints/stage1
python training/sft.py --base Qwen/Qwen3-4B --data ../FaithOR-Supervision/data/sft_train.jsonl --out checkpoints/stage1 --merge-only
python training/train.py --config configs/training.json --stage sft2 --base checkpoints/stage1/merged \
  --data-root ../FaithOR-Supervision --out checkpoints/stage2
python training/sft.py --base checkpoints/stage1/merged --data ../FaithOR-Supervision/data/sft_train.jsonl --out checkpoints/stage2 --merge-only
python training/train.py --config configs/training.json --stage grpo --base checkpoints/stage2/merged \
  --data-root ../FaithOR-Supervision --out checkpoints/grpo      # verl environment; convert global_step_50 with verl.model_merger

Environment. NVIDIA RTX PRO 5000 Blackwell GPUs (48 GB), CUDA 13.2. SFT stages: one GPU each, PyTorch 2.14.1, Transformers 5.18.0, PEFT 0.21.2, bf16 frozen base, peak allocated memory 13.7 GB (stage 1) and 10.3 GB (stage 2); stage 1 processed 29.4 M tokens, stage 2 7.3 M tokens. GRPO: two GPUs, verl 0.9.1, vLLM 0.30.0, PyTorch 2.13.0, about 210 s per step.

Terms

Base model: Qwen/Qwen3-4B (Apache 2.0). Training-data terms are described in the dataset card.

Downloads last month: 25

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for QMJWang/FaithOR-Detector-4B

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Adapter

(1205)

this model

Adapters

1 model

QMJWang
/

FaithOR-Detector-4B