111 lines
4.4 KiB
Markdown
111 lines
4.4 KiB
Markdown
---
|
|
library_name: transformers
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
base_model:
|
|
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- text-generation-inference
|
|
- R1
|
|
- RL
|
|
- Code
|
|
- Math
|
|
---
|
|
|
|
|
|
|
|
# **Pisces-QwenR1-1.5B**
|
|
|
|
> **Pisces-QwenR1-1.5B** is a small reasoning model that enhances the reasoning capabilities of **edge large language models (LLMs)** using **reinforcement learning (RL)**. Fine-tuned from **DeepSeek-R1-Distilled-Qwen-1.5B**, it offers lightweight yet powerful performance in **mathematical reasoning**, **coding**, and **error correction**, making it ideal for edge deployments and on-device intelligent agents.
|
|
|
|
## **Key Improvements**
|
|
1. **Mathematical Reasoning Enhancements**:
|
|
Equipped with refined capabilities in mathematical logic, symbolic computation, step-by-step problem-solving, and numerical accuracy — even in resource-constrained environments.
|
|
|
|
2. **Coding and Debugging Proficiency**:
|
|
Capable of generating, understanding, and debugging code in Python, JavaScript, C++, and other languages, making it a versatile assistant for lightweight coding tasks and educational tools.
|
|
|
|
3. **Intelligent Error Correction**:
|
|
Can identify logical inconsistencies, detect structural errors (in formats like JSON, XML), and offer corrective suggestions — optimized for fast inference and low-latency feedback.
|
|
|
|
4. **Efficient Instruction Following**:
|
|
Fine-tuned to accurately follow multi-step and nested instructions, delivering reliable outputs across compact prompts and conversations.
|
|
|
|
5. **Edge-Optimized Context Handling**:
|
|
Supports long-context inputs up to **128K tokens** and outputs up to **8K tokens**, balancing context-awareness with memory efficiency for edge devices and embedded systems.
|
|
|
|
## **Quickstart with Transformers**
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_name = "prithivMLmods/Pisces-QwenR1-1.5B"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
torch_dtype="auto",
|
|
device_map="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
prompt = "Explain the difference between breadth-first search and depth-first search with Python code examples."
|
|
messages = [
|
|
{"role": "system", "content": "You are a knowledgeable assistant skilled in reasoning, coding, and explanation."},
|
|
{"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. **Edge Inference and Reasoning**:
|
|
Ideal for reasoning and structured output generation on edge devices such as mobile phones, embedded systems, and low-power AI modules.
|
|
|
|
2. **Compact Programming Assistant**:
|
|
Efficient for lightweight coding tasks, debugging, and educational environments where smaller models are preferred.
|
|
|
|
3. **Mathematical Toolkits**:
|
|
Solves mathematical problems and logical reasoning challenges with minimal resource overhead.
|
|
|
|
4. **Conversational Agents**:
|
|
Enables intelligent, context-aware bots and virtual assistants in constrained hardware setups.
|
|
|
|
5. **Multilingual Support & Translation**:
|
|
Useful for lightweight multilingual inference and content generation across various languages.
|
|
|
|
6. **Structured Content Generation**:
|
|
Outputs well-formatted data such as JSON, XML, tables, and Markdown — suitable for embedded AI use cases.
|
|
|
|
## **Limitations**
|
|
|
|
1. **Compute Constraints**:
|
|
While optimized for edge use, still requires adequate hardware (e.g., modern GPUs or NPUs) for efficient large-context processing.
|
|
|
|
2. **Knowledge Cutoff**:
|
|
No real-time access to current events or external data beyond its training.
|
|
|
|
3. **Potential Biases**:
|
|
May exhibit inherited biases or inaccuracies from training data.
|
|
|
|
4. **Variability in Creative Output**:
|
|
Creative writing or abstract tasks may yield variable consistency or style.
|
|
|
|
5. **Prompt Sensitivity**:
|
|
Responses depend heavily on how well prompts are structured — minor changes can impact output significantly. |