76 lines
4.0 KiB
Markdown
76 lines
4.0 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
- zh
|
|||
|
|
base_model:
|
|||
|
|
- HuggingFaceTB/SmolLM2-360M-Instruct
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library_name: transformers
|
|||
|
|
tags:
|
|||
|
|
- Grpo
|
|||
|
|
- text-generation-inference
|
|||
|
|
- Llama
|
|||
|
|
- trl
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
|
|||
|
|
# **SmolLM2-360M-Grpo-r999**
|
|||
|
|
|
|||
|
|
SmolLM2-360M-Grpo-r999 is fine-tuned based on **SmolLM2-360M-Instruct**. SmolLM2 demonstrates significant advances over its predecessor, SmolLM1, particularly in instruction following, knowledge, and reasoning. The **360M** model was trained on **2 trillion tokens** using a diverse combination of datasets: **FineWeb-Edu, DCLM, The Stack**, along with new filtered datasets that we curated and will release soon. We developed the instruct version through **supervised fine-tuning (SFT)** using a combination of public datasets and our own curated datasets.
|
|||
|
|
|
|||
|
|
### **How to Use**
|
|||
|
|
|
|||
|
|
### Transformers
|
|||
|
|
```bash
|
|||
|
|
pip install transformers
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
checkpoint = "prithivMLmods/SmolLM2-360M-Grpo-r999"
|
|||
|
|
|
|||
|
|
device = "cuda" # for GPU usage or "cpu" for CPU usage
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
|
|||
|
|
# for multiple GPUs install accelerate and do `model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")`
|
|||
|
|
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=50, temperature=0.2, top_p=0.9, do_sample=True)
|
|||
|
|
print(tokenizer.decode(outputs[0]))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### **Limitations of SmolLM2-360M-Grpo-r999**
|
|||
|
|
|
|||
|
|
1. **Model Size**: While **360M parameters** provide enhanced capabilities, the model still has limitations in handling highly complex reasoning tasks or long-context dependencies compared to larger models.
|
|||
|
|
|
|||
|
|
2. **Bias and Inaccuracy**: Despite fine-tuning on diverse datasets, the model may generate biased, inaccurate, or factually incorrect responses, particularly for niche topics or specialized knowledge areas.
|
|||
|
|
|
|||
|
|
3. **Context Length**: The model might struggle with very long conversations or extended prompts, potentially leading to truncation or loss of contextual coherence.
|
|||
|
|
|
|||
|
|
4. **Fine-Tuning Specificity**: Performance on specialized domains may require additional fine-tuning with domain-specific datasets.
|
|||
|
|
|
|||
|
|
5. **Generalization**: The model may not generalize as effectively to **rare queries** or **unseen tasks** compared to larger models, sometimes providing generic or incomplete answers.
|
|||
|
|
|
|||
|
|
6. **Limited Multi-Turn Conversations**: While it supports multi-turn interactions, its ability to retain and use context over extended conversations is **not as strong as larger models**.
|
|||
|
|
|
|||
|
|
### **Intended Use of SmolLM2-360M-Grpo-r999**
|
|||
|
|
|
|||
|
|
1. **General-purpose Conversational AI** – Ideal for chatbots, virtual assistants, and interactive applications requiring basic reasoning and knowledge retrieval.
|
|||
|
|
|
|||
|
|
2. **Education & Tutoring** – Supports answering educational queries, explaining concepts, and aiding learning across multiple domains.
|
|||
|
|
|
|||
|
|
3. **Content Generation** – Can generate short-form text, summaries, and brainstorming ideas for writing assistants or creativity tools.
|
|||
|
|
|
|||
|
|
4. **Code Assistance** – Fine-tuned on programming datasets, making it useful for debugging, explaining code, and assisting developers.
|
|||
|
|
|
|||
|
|
5. **Instruction Following** – Optimized for following structured commands, making it suitable for task-based applications.
|
|||
|
|
|
|||
|
|
6. **Prototyping & Experimentation** – Lightweight model for **fast deployment** in new AI applications, balancing performance with efficiency.
|
|||
|
|
|
|||
|
|
7. **Low-Resource Environments** – Runs on **edge devices, mobile apps, and local servers** where larger models are infeasible.
|
|||
|
|
|
|||
|
|
8. **Research & Development** – Can be used as a base model for **further fine-tuning** or model optimizations.
|