112 lines
4.3 KiB
Markdown
112 lines
4.3 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
datasets:
|
||
|
|
- simplescaling/aime24_figures
|
||
|
|
- amphora/QwQ-LongCoT-130K
|
||
|
|
- HuggingFaceH4/MATH-500
|
||
|
|
- RyotaKadoya1993/math-5000-nemotron-v2
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen2-1.5B-Instruct
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- text-generation-inference
|
||
|
|
- thinker
|
||
|
|
- math
|
||
|
|
---
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
# **Open-Xi-Math-Preview**
|
||
|
|
|
||
|
|
> **Open-Xi-Math-Preview** is a **mathematics-focused reasoning model** fine-tuned on **Qwen2-1.5B-Instruct**, utilizing a **modular dataset** designed for enhancing **mathematical thinking**. It provides robust capabilities in symbolic reasoning, structured deduction, and compact coding — optimized for edge deployment on **resource-constrained devices**.
|
||
|
|
|
||
|
|
## **Key Improvements**
|
||
|
|
|
||
|
|
1. **Mathematical Reasoning via Modular Data**:
|
||
|
|
Fine-tuned on diverse and structured math-focused datasets to handle problem-solving, symbolic computation, and multi-step derivations with efficiency on low-power devices.
|
||
|
|
|
||
|
|
2. **Compact Coding & Math Assistant**:
|
||
|
|
Understands multiple programming languages and math representations (e.g., LaTeX, symbolic algebra). Ideal for math-enhanced embedded coding and problem-solving environments.
|
||
|
|
|
||
|
|
3. **Error Detection in Structured Data**:
|
||
|
|
Accurately detects and corrects logical errors, malformed math expressions, and data structures (e.g., JSON, XML, LaTeX), all while maintaining low inference latency.
|
||
|
|
|
||
|
|
4. **Instruction Following for Problem-Solving**:
|
||
|
|
Enhanced with strong instruction-following performance, particularly for step-wise solutions in math word problems, logic puzzles, and equation derivations.
|
||
|
|
|
||
|
|
5. **Extended Context Support**:
|
||
|
|
Supports **128K token inputs** and **8K token outputs**, enabling it to work with long math chains-of-thought and proofs, while remaining lightweight enough for edge inference.
|
||
|
|
|
||
|
|
## **Quickstart with Transformers**
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model_name = "your-username/Open-Xi-Math-Preview"
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
model_name,
|
||
|
|
torch_dtype="auto",
|
||
|
|
device_map="auto"
|
||
|
|
)
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
|
|
||
|
|
prompt = "Solve the equation: 2x^2 - 4x - 6 = 0. Show all steps."
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": "You are a helpful and concise mathematical reasoning assistant."},
|
||
|
|
{"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**
|
||
|
|
|
||
|
|
1. **Math-Centric Edge Applications**:
|
||
|
|
Designed for embedded AI systems in calculators, educational tools, and mobile math tutoring.
|
||
|
|
|
||
|
|
2. **Advanced Math Reasoning**:
|
||
|
|
Effective for solving algebra, geometry, calculus, and competition math problems using logical derivation.
|
||
|
|
|
||
|
|
3. **Educational & Instructional Aids**:
|
||
|
|
Useful for step-by-step teaching in math-heavy domains like STEM education, coding classes, and robotics kits.
|
||
|
|
|
||
|
|
4. **Low-Latency Math Agents**:
|
||
|
|
Deployable in customer support bots, interactive kiosks, and STEM-based IoT systems for fast math-based interactions.
|
||
|
|
|
||
|
|
5. **Structured Output Generation**:
|
||
|
|
Generates LaTeX, JSON, or tabular formats for math answers and reasoning in structured pipelines.
|
||
|
|
|
||
|
|
## **Limitations**
|
||
|
|
|
||
|
|
1. **Edge Hardware Still Required**:
|
||
|
|
Though lightweight, best used with devices equipped with NPUs, GPUs, or optimized ML accelerators.
|
||
|
|
|
||
|
|
2. **No Internet or Real-Time Info**:
|
||
|
|
Static knowledge cutoff; cannot retrieve or interact with live external data sources.
|
||
|
|
|
||
|
|
3. **Not Suited for Creative Tasks**:
|
||
|
|
Focused on deterministic reasoning — not built for abstract, poetic, or generative creative writing.
|
||
|
|
|
||
|
|
4. **Prompt Sensitivity**:
|
||
|
|
Clear, structured prompts yield more accurate reasoning; ambiguous questions may degrade output quality.
|
||
|
|
|
||
|
|
5. **Potential Dataset Biases**:
|
||
|
|
Model may carry forward biases or inconsistencies present in the training datasets; vet outputs in critical settings.
|