README.md

---
license: apache-2.0
language:
- en
tags:
- math
- reasoning
- text-generation
- research-preview
- small-language-models
- sft
- architecture-research
- sonamath
pipeline_tag: text-generation
---

# SonaMath-0.5B

> ## Important (please read)
>
> **The weights in this repository are from the SFT stage only.**
>
> - They are **not** trained with our experimental RL method.
> - They are **not** a drop-in replacement for GRPO / group-based optimizers.
> - **GSM8K numbers below measure this SFT checkpoint**, not an RL algorithm.
>
> We **also** research a new RL formulation (intended as an alternative to GRPO-style group optimization) and a custom reasoning architecture for small LMs. That work is **active / unfinished** and will be reported **separately** when we have matched-compute experiments and RL checkpoints.
>
> This page is a **transparent research preview** under limited compute — not a claim that “new RL already beats GRPO.”

---

**SonaMath** is a **research preview** of a **custom architecture** for **math reasoning in the small-language-model (SLM) regime** (**approx. 0.5B** parameters), trained under a **strict compute budget** (**approx. 3B** pretrain tokens + **light SFT**).

> Formerly previewed as *JunMath*; the project brand is now **SonaMath**.

## What this release is / is not

| | |
|---|---|
| **Is** | Early **SFT** weights + tokenizer + public eval protocol |
| **Is** | Evidence that a **small custom model** can start learning math format/reasoning under few tokens |
| **Is** | A call for compute to scale data, ablations, and (later) RL experiments |
| **Is not** | An RL-trained model |
| **Is not** | A finished “GRPO killer” or published RL baseline win |
| **Is not** | A drop-in `transformers` `AutoModel` checkpoint (custom runtime) |

## How to use (inference)

> **Important:** SonaMath uses a **custom architecture**.  
> It does **not** load with `AutoModelForCausalLM.from_pretrained(...)` / standard Hugging Face `pipeline` yet.  
> Public files are **weights + tokenizer**; the research runtime loader is required to run generation.

### 1) Download files from this repo

| File | Role |
|------|------|
| `sonamath.pt` | SFT weights package |
| `tokenizer.json` | BPE tokenizer (32k) |
| `tokenizer_config.json` / `special_tokens_map.json` | Tokenizer metadata |
| `config.json` | Public model card config |

```bash
# Option A: huggingface-cli
huggingface-cli download HaniAI/SonaMath-0.5B --local-dir ./SonaMath-0.5B

# Option B: Python
pip install -U "huggingface_hub>=0.23"
```

```python
from huggingface_hub import snapshot_download

path = snapshot_download("HaniAI/SonaMath-0.5B", local_dir="SonaMath-0.5B")
print("downloaded to", path)
# weights: SonaMath-0.5B/sonamath.pt
# tokenizer: SonaMath-0.5B/tokenizer.json
```

Downloading the repo **does count as usage** and is the supported way to fetch artifacts while the full open runtime is still under development.

### 2) Prompt format (required)

The model is trained with explicit problem / thinking tags. Build prompts like:

```text
<problem>
{your grade-school math word problem here}
</problem>
<think>
```

The model continues with reasoning and typically closes with:

```text
... reasoning ...
<answer>
{final number}
</answer>
```

### 3) Recommended decoding (matches public GSM8K number)

| Setting | Value |
|---------|-------|
| Mode | **greedy** (`temperature = 0`) |
| Max new tokens | **1024** |
| Stop | `</answer>` or EOS when available in your runtime |

### 4) Pseudocode (custom research runtime)

```python
# Pseudocode — requires the SonaMath research runtime (not a stock transformers model).
# Full open loader will be linked here when released.

from sonamath_runtime import load_sonamath, generate  # research package (not on PyPI yet)

model, tokenizer = load_sonamath(
    hub_id="HaniAI/SonaMath-0.5B",   # or local folder from snapshot_download
    device="cuda",
)

problem = (
    "Natalia sold clips to 48 of her friends in April, and then she sold "
    "half as many clips in May. How many clips did Natalia sell altogether "
    "in April and May?"
)

text = generate(
    model,
    tokenizer,
    problem=problem,
    max_new_tokens=1024,
    temperature=0.0,  # greedy — same as public eval
)
print(text)
```

### 5) What works today vs later

| Today (this card) | Coming later |
|-------------------|--------------|
| Download **weights + tokenizer** from Hub | Public minimal inference package / Space |
| Reproduce **eval protocol** (greedy, max_new=1024) | Drop-in scripts without private stack |
| Interactive testing in the author’s research stack | Optional HF Space demo |

If you only need artifacts for inspection, fine-tuning research, or offline study:

```python
from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("HaniAI/SonaMath-0.5B", "sonamath.pt")
tok  = hf_hub_download("HaniAI/SonaMath-0.5B", "tokenizer.json")
print(ckpt, tok)
```

### 6) Not supported (yet)

```python
# ❌ Will NOT work — not a standard transformers architecture export
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("HaniAI/SonaMath-0.5B")  # no
```

Questions / runtime access requests: open a **Discussion** on this model page.

## Two research tracks (keep them separate)

### Track A — Architecture + data-efficient SFT *(this repo)*

- Custom architecture specialized for **multi-step math reasoning** in **small LMs**
- Pretrain **approx. 3B** tokens (math-focused, English-centric)
- **Light SFT** on math Q&A / solution-style data
- **Public artifact:** `sonamath.pt` (SFT-stage, weights only)

### Track B — RL theory & methods *(research in progress; not this checkpoint)*

- Studying a **new RL formulation** as an **alternative to group-based / GRPO-style** post-training
- Status: **theory + early experiments** — **no claim of superiority** on this model card
- **Will not** use Track A’s GSM8K score as “proof” of Track B

> **SFT checkpoint for architecture research now; RL-vs-GRPO is future work with separate evals.**

## Evaluation (SFT checkpoint only)

### Protocol

| Item | Setting |
|------|---------|
| Benchmark | **GSM8K test (full)** |
| Samples | **n = 1319** (entire official test split) |
| Decoding | **greedy** (`temperature = 0`) |
| Max new tokens | **1024** |
| Metric | numeric / exact match on final answer |
| Stage | **SFT only** |

### Results

| Model | Stage | Params (approx.) | GSM8K full (greedy) | Parse rate | 95% Wilson CI |
|------:|-------|-----------------:|--------------------:|-----------:|--------------:|
| **SonaMath-0.5B (this repo)** | **SFT** | **0.5B** | **13.8% (182/1319)** | **99.3%** | **[12.0%, 15.8%]** |

**Notes**

- This is the **authoritative public number** for this release (full test set).  
- Absolute accuracy is **well below** large or math-specialized models — expected under **approx. 3B** pretrain tokens.  
- When Track B (RL) is ready, we will report **SFT vs GRPO-style vs our RL** under **matched compute**, on **separate checkpoints**.

## Highlights

| | |
|---|---|
| **Brand** | **SonaMath** |
| **Parameters** | approx. 0.5B |
| **Released stage** | **SFT preview** |
| **Pretrain tokens** | approx. 3B |
| **Architecture** | Custom (details partially private while under development) |
| **RL (Track B)** | In progress — **not** applied to these weights |
| **Language** | English math focus |
| **GSM8K (full)** | **13.8% (182/1319)**, greedy, max_new=1024 |

## Files

| File | Description |
|------|-------------|
| `sonamath.pt` | **SFT** weights-only package (no optimizer) |
| `config.json` | Public metadata |
| `tokenizer.json` (+ config maps) | 32k BPE, digit-aware |

## Intended use

- Research discussion on **small reasoning LMs** and **data-efficient** math training  
- Download weights/tokenizer for inspection and future runtime integration  
- Planning compute for larger pretrain, cleaner SFT, and **future** RL ablations  

**Not intended for:** production tutoring, grading, high-stakes decisions, or citing this repo as an RL result.

## Limitations

- Low absolute GSM8K accuracy vs large / math-specialized models  
- May hallucinate arithmetic and multi-step logic — **verify answers**  
- Custom architecture: full inference stack not fully open as `transformers` yet  
- **No public RL checkpoint** in this release  

## Compute context & GPU ask

| Stage | Status / budget |
|------|------------------|
| Pretrain | approx. 3B tokens (done, limited hardware) |
| SFT | Light schedule → **this release** |
| RL (vs GRPO-style) | Research track — needs more GPU for fair ablations |

Open a **Discussion** on this page if you can support **A100/H100 hours** for scale-up and matched-compute RL ablations.

## Safety

Outputs can be confidently wrong. Do not use without human checking for education, finance, or safety-critical settings.

## Citation

```bibtex
@misc{sonamath2026sft,
  title  = {SonaMath-0.5B: SFT-Stage Research Preview of a Small Math Reasoning Model},
  author = {HaniAI},
  year   = {2026},
  note   = {SFT weights only; GSM8K full test 13.8\\% (182/1319), greedy, max\\_new=1024; RL research is separate},
  url    = {https://huggingface.co/HaniAI/SonaMath-0.5B}
}
```

## License

Apache-2.0 (weights and tokenizer files in this repository).