CARE-Qwen2.5-7B-Instruct is a 7.61B parameter instruction-tuned language model based on Qwen2.5-7B-Instruct, enhanced with native retrieval-augmented reasoning capabilities through the CARE (Context-Aware Retrieval-Enhanced reasoning) framework. This model has been specifically trained to improve context fidelity and reduce hallucinations by teaching the model to explicitly integrate in-context evidence within its reasoning process.
Key Features:
Native retrieval-augmented reasoning: Dynamically identifies and incorporates relevant evidence from input context
Improved context fidelity: Significantly better adherence to provided context, especially when it contradicts parametric knowledge
Enhanced multi-hop reasoning: Superior performance on complex reasoning tasks requiring evidence integration
Structured reasoning output: Generates reasoning chains with explicit evidence citations using <think> and <retrieval> tags
Model Details
Model Type: Causal Language Model (Enhanced with Retrieval-Augmented Reasoning)
Base Model: Qwen2.5-7B-Instruct
Parameters: 7.61B total, 6.53B non-embedding
Architecture: Transformer with RoPE, SwiGLU, RMSNorm, and Attention QKV bias
Training Framework: Two-phase training (SFT + Reinforcement Learning with GRPO)
Training Process
The model was trained using a novel two-phase approach:
Phase 1 - Supervised Fine-Tuning (SFT):
Dataset: 7,739 instances from HotpotQA with retrieval-augmented reasoning chains
Purpose: Establish evidence integration patterns and reasoning format
Training: 3 epochs with LoRA (r=8, α=16), AdamW optimizer
Phase 2 - Reinforcement Learning:
Method: Group Relative Policy Optimization (GRPO)
Curriculum Learning: Gradual transition from DROP (easy) to MS MARCO (hard)
Rewards: Accuracy + Format + Retrieval consistency
Training: 350 steps with multi-aspect reward optimization
System Prompt
The model uses an enhanced system prompt that enables structured reasoning with evidence retrieval:
You are Qwen, created by Alibaba Cloud. You are a helpful assistant. You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. WITHIN the thinking process, make reference to the relevant texts in the prompt that provide critical information to move the reasoning process forward. The referenced texts MUST BE enclosed within <retrieval> </retrieval> tags, and MUST BE placed within the reasoning process only. The final answer MUST BE put at the end of the response after "Answer:".
Note: This system prompt is automatically applied when using the default chat template.
Usage
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model_name ="sheryc/Qwen2.5-7B-Instruct-CARE"45model = AutoModelForCausalLM.from_pretrained(6 model_name,7 torch_dtype=torch.float16,8 device_map="auto"9)10tokenizer = AutoTokenizer.from_pretrained(model_name)1112# Example usage13context ="""John went to the movies with his mom last week. They watched the latest superhero movie, which was quite popular. The ticket price was $15. According to the local cinema's website, ticket prices range from $10 to $12 for regular screenings and from $13 to $16 for special releases."""1415question ="Was the ticket price John's mom paid for the movie reasonable?"1617messages =[18{"role":"user","content":f"{question}\n\nContext:{context}"}19]2021tokenized_chat = tokenizer.apply_chat_template(22 messages,23 tokenize=True,24 add_generation_prompt=True,25 return_tensors="pt"26)2728generated_ids = model.generate(tokenized_chat.to(model.device), max_new_tokens=512)29output_text = tokenizer.decode(outputs[0])
Expected Output Format:
<think>
The context states John watched the latest superhero movie. <retrieval>The ticket price was $15.</retrieval> The context provides price ranges: <retrieval>ticket prices range from $10 to $12 for regular screenings and from $13 to $16 for special releases.</retrieval> Since this was a popular latest superhero movie, it likely qualifies as a special release. Therefore, the $15 price falls within the $13-$16 range for special releases.
</think>
Answer: Yes, the ticket price was reasonable.
Training Data
SFT Phase: HotpotQA with labeled supporting facts (7,739 instances)
RL Phase:
DROP dataset (77,409 training instances) - Easy curriculum phase
MS MARCO - Hard curriculum phase
Evaluation: LongBench, CofCA, and other QA benchmarks
License
Please refer to the original Qwen2.5 license terms.
Citation
bibtex
1@inproceedings{wang2025care,
2 title={Improving Context Fidelity via Native Retrieval-Augmented Reasoning},
3 author={Wang, Suyuchen and Wang, Jinlin and Wang, Xinyu and Li, Shiqi and Tang, Xiangru and Hong, Sirui and Chang, Xiao-Wen and Wu, Chenglin and Liu, Bang},
4 booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
5 year={2025}
6}
78@misc{qwen2.5,
9 title = {Qwen2.5: A Party of Foundation Models},
10 url = {https://qwenlm.github.io/blog/qwen2.5/},
11 author = {Qwen Team},
12 month = {September},
13 year = {2024}
14}
Contact
For questions about the model or to report issues, please visit the CARE project homepage or contact the authors.
