---
license: apache-2.0
tags:
  - nebula-s
  - svms
  - math-reasoning
  - competition-math
library_name: transformers
---

# Nebula-S-v1

**Nebula-S-v1** is a reasoning-enhanced language model using the **SVMS (Structured-Vector Multi-Stream)** architecture.

## Architecture

SVMS adds a multi-stream reasoning layer on top of a frozen 4B-parameter backbone:

- **Structured Consistency**: Topological constraint forcing cross-stream coherence
- **Learned Router**: Per-token stream weighting
- **Delta Logits**: Learnable correction to backbone predictions

| Component | Details |
|---|---|
| **Trainable** | **400M** / 4.4B total |

## Quick Start


```bash
pip install torch transformers huggingface-hub
```
Option 1: Using huggingface_hub
```
from huggingface_hub import snapshot_download
import sys



snapshot_download("decompute/Nebula-S-v1", local_dir="./Nebula-S-v1")
sys.path.insert(0, "./Nebula-S-v1")
from nebula_s import load_nebula_s

model, tokenizer = load_nebula_s("./Nebula-S-v1", device="cuda")
```
Option 2: Using git clone
```
git lfs install
git clone https://huggingface.co/punitdecomp/Nebula-S-v1

import sys
sys.path.insert(0, "./Nebula-S-v1")
from nebula_s import load_nebula_s

model, tokenizer = load_nebula_s("./Nebula-S-v1", device="cuda")
```
Generate a response
```
messages = [{"role": "user", "content": "Solve step by step: what is 17 * 23?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
response = model.generate(
    inputs["input_ids"], inputs["attention_mask"],
    tokenizer, max_new_tokens=2048, temperature=0.7
)
print(response)
```

## Training

- **Data**: Orca Math Word Problems (200K)
- **Steps**: 1000
- **Method**: Adapter-only (backbone frozen)

## Evaluation Results

Nebula-S-v1 was evaluated on several challenging benchmarks:

| Benchmark | Score |
|---|---|
| GSM8K | 90% |
| GPQA | 70.5% |
| HMMT (November 2025) | 67% |
| MMLU-Pro | 79.7% |

These results demonstrate strong performance for a 4B-parameter model, particularly on math reasoning (GSM8K) and advanced knowledge/competition-level tasks.

## License

Apache 2.0. Backbone derived from an Apache-2.0 licensed base model.