107 lines
3.3 KiB
Markdown
107 lines
3.3 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
datasets:
|
||
|
|
- sequelbox/Celestia3-DeepSeek-R1-0528
|
||
|
|
base_model:
|
||
|
|
- HuggingFaceTB/SmolLM2-135M-Instruct
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- trl
|
||
|
|
- text-generation-inference
|
||
|
|
- re-think
|
||
|
|
- reasoning
|
||
|
|
---
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
# **SmolLM2-Rethink-135M**
|
||
|
|
|
||
|
|
> **SmolLM2-Rethink-135M** is an experimental lightweight model trained on the **Celestia3-DeepSeek-R1-0528** reasoning dataset. Based on the **SmolLM2-135M-Instruct** architecture, this model is specifically optimized for reasoning, structured outputs, and efficient small-scale deployment. Despite its compact size (135M parameters), it demonstrates strong capabilities in logical deduction, conversational coherence, and lightweight inference tasks.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Key Highlights**
|
||
|
|
|
||
|
|
1. **Compact & Efficient**
|
||
|
|
Lightweight architecture (135M) suitable for fast inference, mobile applications, and edge deployment.
|
||
|
|
|
||
|
|
2. **Reasoning-Centric Training**
|
||
|
|
Fine-tuned on high-quality reasoning and instruction datasets like **Celestia3-DeepSeek-R1-0528**, focusing on multi-step logical thinking.
|
||
|
|
|
||
|
|
3. **Low-Resource Optimization**
|
||
|
|
Designed to run effectively on CPUs or single-GPU setups with minimal memory footprint.
|
||
|
|
|
||
|
|
4. **Structured Outputs**
|
||
|
|
Supports generation of clean, structured content including lists, steps, tables, and JSON-like responses.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Quickstart with 🤗 Transformers**
|
||
|
|
|
||
|
|
```python
|
||
|
|
%%capture
|
||
|
|
!pip install transformers
|
||
|
|
```
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
checkpoint = "prithivMLmods/SmolLM2-Rethink-135M"
|
||
|
|
device = "cuda" # or "cpu"
|
||
|
|
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
|
||
|
|
|
||
|
|
messages = [{"role": "user", "content": "What is gravity?"}]
|
||
|
|
input_text = tokenizer.apply_chat_template(messages, tokenize=False)
|
||
|
|
print(input_text)
|
||
|
|
|
||
|
|
inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
|
||
|
|
outputs = model.generate(
|
||
|
|
inputs,
|
||
|
|
max_new_tokens=1024,
|
||
|
|
temperature=0.2,
|
||
|
|
top_p=0.9,
|
||
|
|
do_sample=True
|
||
|
|
)
|
||
|
|
|
||
|
|
print(tokenizer.decode(outputs[0]))
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Intended Use**
|
||
|
|
|
||
|
|
* **Instruction Following & QA**
|
||
|
|
Good for answering simple questions, following short instructions, and general user interactions.
|
||
|
|
|
||
|
|
* **Educational Tools**
|
||
|
|
Suitable for lightweight tutoring bots or classroom assistants on low-compute setups.
|
||
|
|
|
||
|
|
* **Reasoning Tasks**
|
||
|
|
Performs well on logic puzzles, multi-step reasoning, and chain-of-thought queries.
|
||
|
|
|
||
|
|
* **Prototype Agents & Microservices**
|
||
|
|
Can be deployed in memory-efficient environments or as modular AI components.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Limitations**
|
||
|
|
|
||
|
|
1. **Limited Knowledge Capacity**
|
||
|
|
Due to small parameter size, lacks the depth and breadth of large-scale models.
|
||
|
|
|
||
|
|
2. **Short-Term Context Handling**
|
||
|
|
Performs best with short to moderate-length prompts; lacks extended context support.
|
||
|
|
|
||
|
|
3. **Creative Generation Limitations**
|
||
|
|
Output may lack diversity or depth in open-ended storytelling or imaginative tasks.
|
||
|
|
|
||
|
|
4. **Token Budget**
|
||
|
|
Smaller output range; optimized for shorter and structured completions.
|
||
|
|
|
||
|
|
5. **Basic Multilingual Support**
|
||
|
|
Some support for multilingual input, but less accurate than larger multilingual models.
|