# SweetFlippy Coder - Private Deployment

## Model Information
- **Base Model**: Qwen/Qwen2.5-Coder-32B-Instruct
- **Training**: SFT on 19,133 examples
- **Checkpoint**: checkpoint-400
- **Evaluation**: 100% pass rate (5/5 tasks)

## Deployment Locations

### 1. Local Merged Model
**Path**: `/data/outputs/sweetflippy-agent/merged-model/`
- **Size**: 62GB
- **Format**: HuggingFace (14 shards)
- **Status**: ✅ Ready for inference

### 2. Backblaze B2 (Private)
**Location**: `b2:SweetflippyAICreated/models/sweetflippy-coder-merged/`
- **Status**: Uploading...
- **Access**: Private bucket
- **Credentials**: Configured in `.env`

### 3. HuggingFace Hub (Pending)
**Repo**: `Sweetflips/qwen2.5-coder-32b-agent-sft-private`
- **Status**: Token needs refresh
- **Privacy**: Private repository

## Usage

### Load from Local
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "/data/outputs/sweetflippy-agent/merged-model",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(
    "/data/outputs/sweetflippy-agent/merged-model",
    trust_remote_code=True
)
```

### Download from B2
```bash
rclone sync b2:SweetflippyAICreated/models/sweetflippy-coder-merged ./local-model/
```

## Model Performance
- **Training Loss**: 0.4106 (final)
- **Token Accuracy**: 86.94%
- **Evaluation**: 100% pass rate
- **Processed Tokens**: ~78M during training

## Training Details
- **Epochs**: 1.0/3 (400 steps)
- **Batch Size**: 128 effective
- **Learning Rate**: Cosine decay (1.5e-4 → 5.15e-7)
- **Hardware**: 8× H200 GPUs
- **Training Time**: 4h 55m