README.md

---
language:
- en
- multilingual
license: apache-2.0
library_name: transformers
tags:
- qwen
- qwen3.5
- finetuned
- astrophysics
- science
- cot
- chain-of-thought
- unsloth
- lora
- llama.cpp
- gguf
base_model: Qwen/Qwen3.5-0.8B
---

# Qwen3.5-0.8B-Astro-Reasoning-v1

This is a finetuned version of [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) specialized for **astrophysics problem-solving** and **chain-of-thought reasoning**.

## Model Description

- **Base Model:** Qwen/Qwen3.5-0.8B
- **Model Size:** 0.8B parameters
- **Architecture:** Causal Language Model with Vision Encoder
- **Context Length:** 1,024 tokens (training), up to 262,144 tokens (inference)
- **Training Method:** LoRA (Low-Rank Adaptation)
- **Precision:** BF16 training, F16 inference (GGUF)

## Training Details

### Hardware
- **GPU:** NVIDIA GeForce RTX 3060 (12GB VRAM)
- **Training Framework:** Unsloth (4-bit quantization)
- **Training Time:** ~32 minutes
- **Effective Batch Size:** 8 (batch_size=1, gradient_accumulation=8)

### Hyperparameters
| Parameter | Value |
|-----------|-------|
| LoRA Rank (r) | 8 |
| LoRA Alpha | 8 |
| Learning Rate | 2e-4 |
| Max Steps | 300 |
| Warmup Steps | 10 |
| Sequence Length | 1,024 |
| Optimizer | adamw_8bit |
| Weight Decay | 0.01 |

### Training Results
- **Final Loss:** 1.656
- **Loss Reduction:** 14% (from 1.924 to 1.656)
- **Epochs:** 0.22

## Dataset

The model was finetuned on 12,357 high-quality examples from two sources:

### 1. Gemini-3 Pro Dataset (10,031 examples)
- **Domain:** Astrophysics
- **Difficulty:** Extreme-level problems
- **Content:** Complex astrophysical concepts including:
  - Eddington Luminosity in Porous Atmospheres
  - Electron Capture Supernovae (ECSN)
  - Beta Cephei Pulsations
  - Type Ia Supernova Progenitors
  - Neutrino Oscillations
  - CNO Cycle Branching
  - Gravitational Radiation Reaction
  - And more...

### 2. Distilled Corpus (2,326 examples)
- **Domains:** Mathematics, coding, natural language inference
- **Features:** Chain-of-thought reasoning with detailed solutions
- **Format:** Problem → Thinking → Solution

## Model Capabilities

This model excels at:
- ✅ **Astrophysics problem-solving** with step-by-step reasoning
- ✅ **Complex scientific calculations** and derivations
- ✅ **Chain-of-thought reasoning** for multi-step problems
- ✅ **Mathematical reasoning** with detailed explanations
- ✅ **Technical documentation** and analysis

## Usage

### With llama.cpp (Recommended)

```bash
llama-cli \
    -m qwen3.5-0.8b-astro-reasoning-v1.gguf \
    --chat-template chatml \
    -c 2048 \
    -n 512 \
    --temp 0.7 \
    -cnv
```

### With Python (Transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "firmanda/qwen3.5-0.8b-astro-reasoning-v1"
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("firmanda/qwen3.5-0.8b-astro-reasoning-v1")

prompt = "Solve this astrophysics problem: What is the Eddington luminosity?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### With Unsloth (For Further Training)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="firmanda/qwen3.5-0.8b-astro-reasoning-v1",
    max_seq_length=2048,
    load_in_4bit=True,
)
```

## Example Outputs

### Example 1: Astrophysics Problem
```
User: Calculate the Eddington luminosity for a 10 solar mass star.

Assistant: The Eddington luminosity is given by:

L_Edd = (4πGMc) / κ

Where:
- G = 6.674 × 10^-11 m^3 kg^-1 s^-2
- M = 10 M_☉ = 1.989 × 10^31 kg
- c = 3 × 10^8 m/s
- κ = 0.34 m^2/kg (Thomson scattering for ionized hydrogen)

Calculation:
L_Edd = (4π × 6.674 × 10^-11 × 1.989 × 10^31 × 3 × 10^8) / 0.34
L_Edd ≈ 1.26 × 10^32 W ≈ 3.28 × 10^5 L_☉

This is the maximum luminosity where radiation pressure doesn't overcome gravity.
```

### Example 2: Chain-of-Thought Reasoning
```
User: A class of 12 students shares 108 oranges, but 36 are bad. 
How many fewer oranges per student?

Assistant: Let me solve this step by step:

1. Total oranges: 108
2. Bad oranges: 36
3. Good oranges: 108 - 36 = 72

Original distribution:
- 108 oranges / 12 students = 9 oranges each

New distribution:
- 72 oranges / 12 students = 6 oranges each

Difference: 9 - 6 = 3 oranges fewer per student.
```

## Model Limitations

- **Context Window:** Optimized for 1,024 tokens during training (can handle up to 262,144 for inference)
- **Domain Specificity:** Best performance on astrophysics and scientific reasoning; may underperform on general chat
- **Factual Accuracy:** While trained on scientific content, always verify critical calculations
- **Language:** Primarily trained on English content
- **Reasoning Mode:** Qwen3.5 0.8B operates in non-thinking mode by default

## Evaluation

The model was evaluated on:
- Training loss reduction: **14% improvement**
- Gradient norms remained stable throughout training
- No signs of overfitting observed

### Hardware Compatibility

**Minimum Requirements:**
- **Inference:** 2GB VRAM (F16 GGUF)
- **Training:** 8GB+ VRAM recommended

**Tested On:**
- NVIDIA RTX 3060 12GB (training & inference)

## Files Included

```
qwen3.5-0.8b-astro-reasoning-v1/
├── config.json                    # Model configuration
├── model.safetensors              # Model weights (LoRA adapters)
├── README.md                      # This file
├── qwen3.5-0.8b-astro-reasoning-v1.gguf  # GGUF format for llama.cpp
└── training_info.md               # Detailed training logs
```
## Acknowledgments

- **Base Model:** [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) by Alibaba Cloud Qwen Team
- **Training Framework:** [Unsloth](https://github.com/unslothai/unsloth) for efficient finetuning
- **GGUF Conversion:** [llama.cpp](https://github.com/ggerganov/llama.cpp) for optimized inference

## License

This model is licensed under the Apache 2.0 License, same as the base Qwen3.5 model.

---

**Last Updated:** March 2026  
**Model Version:** v1.0