---
base_model: state-spaces/mamba-790m-hf
library_name: peft
datasets:
- PJMixers-Dev/dolphin-deepseek-1k-think-1k-response-filtered-ShareGPT
---

# Mamba-790M Reasoning Faithfulness Model

Fine-tuned Mamba-790M model for evaluating reasoning trace faithfulness in language models.

## Model Description

This is a LoRA-adapted version of [Mamba-790M](https://huggingface.co/state-spaces/mamba-790m-hf) trained to generate responses that faithfully follow provided reasoning traces.

- **Base Model**: state-spaces/mamba-790m-hf (790M parameters)
- **Adapter Type**: LoRA (Low-Rank Adaptation)
- **Training Data**: [Dolphin-DeepSeek Reasoning Dataset](https://huggingface.co/datasets/PJMixers-Dev/dolphin-deepseek-1k-think-1k-response-filtered-ShareGPT)

## Intended Use

This model is designed for research on:
- **Reasoning faithfulness**: Testing if model outputs align with stated reasoning
- **AI interpretability**: Understanding how models follow (or deviate from) reasoning traces
- **Alignment research**: Measuring consistency between reasoning and conclusions

### Example Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# Load model
base_model = AutoModelForCausalLM.from_pretrained(
    "state-spaces/mamba-790m-hf",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, "NakshJain/mamba-790m-resoning")
tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-790m-hf", trust_remote_code=True)

# Format: <user>question</user><think>reasoning</think><answer>
prompt = "<user>What is 15 + 27?</user><think>Let me add: 15 + 27 = 42</think><answer>"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)
# Output: 42
```

## Training Details

### Training Data

- **Source**: Dolphin-DeepSeek filtered ShareGPT conversations
- **Training Set**: 18500 reasoning examples
- **Validation Set**: 1500 examples
- **Test Set**: 500 examples
- **Format**: Structured as `<user>query</user><think>reasoning</think><answer>response</answer>`

### Training Configuration

- **LoRA Rank (r)**: 32
- **LoRA Alpha**: 32
- **LoRA Dropout**: 0.05
- **Target Modules**: `in_proj`, `x_proj`, `dt_proj`
- **Learning Rate**: 1.5e-5
- **Batch Size**: 4 (effective: 8 with gradient accumulation)
- **Epochs**: 1
- **Optimizer**: AdamW
- **LR Schedule**: Cosine with 3% warmup


## Citation

If you use this model, please cite:
```bibtex
@misc{mamba-reasoning-faithfulness-2024,
  author = {Naksh Jain},
  title = {Mamba-790M Reasoning Faithfulness Model},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/NakshJain/mamba-790m-resoning}
}
```

## Acknowledgments

- **Base Model**: [Mamba](https://github.com/state-spaces/mamba) by Tri Dao and Albert Gu
- **Training Dataset**: Dolphin-DeepSeek filtered by PJMixers-Dev
- **Framework**: HuggingFace Transformers, PEFT

## License

Apache 2.0 (inherits from base Mamba model)