97 lines
3.9 KiB
Markdown
97 lines
3.9 KiB
Markdown
---
|
|
library_name: transformers
|
|
tags:
|
|
- text-generation-inference
|
|
- PRM
|
|
- Code
|
|
- Math
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
base_model:
|
|
- Qwen/Qwen2.5-1.5B-Instruct
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
|
|
|
|
# **Deepthink-1.5B-Open-PRM**
|
|
|
|
> **Deepthink-1.5B-Open-PRM** is a **process-supervised reasoning model** fine-tuned from **Qwen2.5 1.5B** using **Process Reward Models (PRM)**. It excels at **step-by-step mathematical problem solving** in both **English** and **Simplified Chinese**, offering interpretable, logically structured responses for use in **education**, **STEM tutoring**, and **lightweight math agents**.
|
|
|
|
## **Key Features**
|
|
|
|
1. **Process Reward Model Supervision (PRM)**
|
|
Fine-tuned with PRMs to reward high-quality intermediate reasoning steps — fostering step-by-step interpretability, accuracy, and educational transparency.
|
|
|
|
2. **Compact Foundation (Qwen2.5 0.5B)**
|
|
Built upon the highly efficient Qwen2.5 1.5B architecture and scaled up through distillation and reward-based alignment to 1.5B parameters, balancing reasoning quality and deployment efficiency.
|
|
|
|
3. **Bilingual Math Capability**
|
|
Fluent in solving and explaining math problems in both **English** and **Simplified Chinese**, making it ideal for multilingual classrooms and tutoring platforms.
|
|
|
|
4. **Process-Supervised Math Reasoning**
|
|
Trained to reason like a teacher — showing each logical step before delivering an answer. Ideal for learners who need to understand the “how” and “why” behind each solution.
|
|
|
|
5. **Long-Context & Word Problem Reasoning**
|
|
Especially proficient with multi-step arithmetic, word problems, logic puzzles, and middle school to early college-level math.
|
|
|
|
## **Quickstart with Transformers**
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_name = "prithivMLmods/Deepthink-1.5B-Open-PRM"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
torch_dtype="auto",
|
|
device_map="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
prompt = "Solve: A tank can be filled by one pipe in 6 hours and emptied by another in 9 hours. How long will it take to fill the tank if both pipes are opened together?"
|
|
|
|
messages = [
|
|
{"role": "system", "content": "You are a helpful math tutor who explains each step clearly."},
|
|
{"role": "user", "content": prompt}
|
|
]
|
|
|
|
text = tokenizer.apply_chat_template(
|
|
messages,
|
|
tokenize=False,
|
|
add_generation_prompt=True
|
|
)
|
|
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
|
|
|
generated_ids = model.generate(
|
|
**model_inputs,
|
|
max_new_tokens=512
|
|
)
|
|
generated_ids = [
|
|
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
|
|
]
|
|
|
|
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
|
```
|
|
|
|
## **Intended Use**
|
|
|
|
- **Math Education Agents**: Tutors that explain problems step by step, helping users build understanding through reasoning.
|
|
- **Bilingual Learning Platforms**: Apps that teach math in both Chinese and English.
|
|
- **STEM-Oriented Assistants**: Supports early-stage problem solving in science and engineering contexts.
|
|
- **Lightweight LLM Deployments**: Optimized for low-resource environments, from browsers to mobile devices.
|
|
|
|
## **Limitations**
|
|
|
|
1. **Domain Specificity**
|
|
Primarily tuned for math reasoning — performance may degrade on unrelated tasks like creative writing or open dialogue.
|
|
|
|
2. **Model Size Constraint**
|
|
While efficient, 1.5B parameters may struggle with highly abstract or very long multi-domain tasks.
|
|
|
|
3. **PRM Bias Generalization**
|
|
PRM training can bias toward rewardable structures — results should still be reviewed for correctness and completeness.
|
|
|
|
4. **Prompt Structure Sensitivity**
|
|
Well-structured queries yield more accurate and educationally useful outputs. |