104 lines
3.8 KiB
Markdown
104 lines
3.8 KiB
Markdown
---
|
|
datasets:
|
|
- prithivMLmods/Poseidon-Reasoning-5M
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
base_model:
|
|
- Qwen/Qwen3-1.7B
|
|
library_name: transformers
|
|
tags:
|
|
- text-generation-inference
|
|
- moe
|
|
- code
|
|
- science
|
|
- biology
|
|
- chemistry
|
|
- thinking
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
|
|
|
|
# **Poseidon-Reasoning-1.7B**
|
|
|
|
> **Poseidon-Reasoning-1.7B** is a general-purpose, high-efficiency reasoning model fine-tuned on **Qwen3-1.7B** using the **Poseidon-Reasoning-5M** dataset (first 70K entries). Designed for **mathematical, scientific, and code-related reasoning**, this model strikes a balance between structured logic and contextual fluency—ideal for domains demanding symbolic precision and algorithmic thought.
|
|
|
|
> \[!note]
|
|
> GGUF: [https://huggingface.co/prithivMLmods/Poseidon-Reasoning-1.7B-GGUF](https://huggingface.co/prithivMLmods/Poseidon-Reasoning-1.7B-GGUF)
|
|
|
|
## **Key Features**
|
|
|
|
1. **Versatile Reasoning Model**
|
|
Finely tuned for multi-domain reasoning tasks, including mathematics, scientific computation, and code logic—capable of navigating structured problem-solving and analytic workflows.
|
|
|
|
2. **Qwen3-1.7B Foundation**
|
|
Built upon **Qwen3-1.7B**, providing multilingual reasoning capability, efficient token handling, and strong alignment with instruction-following tasks.
|
|
|
|
3. **Powered by Poseidon-Reasoning-5M (70K Sample Subset)**
|
|
Trained on a carefully selected subset of 70K entries from the **Poseidon-Reasoning-5M** dataset—focusing on tasks that emphasize **symbolic accuracy**, **step-by-step thinking**, and **STEM-relevant clarity**.
|
|
|
|
4. **Balanced Thinking Mode**
|
|
Supports structured, guided thinking without excessive hallucination or unnecessary verbosity. Ideal for prompt-driven logic tasks with moderate complexity.
|
|
|
|
5. **Rich Format Output**
|
|
Outputs include **Markdown**, **Python**, **LaTeX**, and tabular structures—helpful for notebooks, scientific documentation, and programmatic outputs.
|
|
|
|
6. **1.7B Parameter Footprint**
|
|
Lightweight enough to run on **mid-tier GPUs or CPU-only environments**, while offering scalable reasoning power for research, teaching, and light automation.
|
|
|
|
## **Quickstart with Transformers**
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_name = "prithivMLmods/Poseidon-Reasoning-1.7B"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
torch_dtype="auto",
|
|
device_map="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
prompt = "Solve: What is the derivative of sin(x) * ln(x)?"
|
|
|
|
messages = [
|
|
{"role": "system", "content": "You are a structured reasoning assistant for math, science, and code."},
|
|
{"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=256
|
|
)
|
|
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]
|
|
print(response)
|
|
```
|
|
|
|
## **Intended Use**
|
|
|
|
* General-purpose symbolic reasoning
|
|
* Math and science tutoring, theorem solving, and computational guidance
|
|
* Structured coding under constraints or STEM-based tasks
|
|
* Lightweight environments where interpretability and precision matter
|
|
* Prompt-driven reasoning with deterministic steps
|
|
|
|
## **Limitations**
|
|
|
|
* Not designed for broad open-domain conversation
|
|
* May underperform on creative writing or emotional expression
|
|
* Best results occur with **clear problem statements and goal-directed prompts**
|
|
* Less suitable for speculative or abstract reasoning without structure |