# SmolLM3-3B LoRA — Tool Calling Fine-Tune

LoRA adapter training to teach SmolLM3-3B to emit native tool-call tokens.

## Critical Training Objective

The base model **does not emit structured tool-call tokens**. When asked to use tools, it writes Python code that *calls* the tool as a function instead of emitting the `startPos`/`endPos` (token IDs 128002/128016) sequences that vLLM's Hermes parser expects. This was verified definitively using a raw token inspector (`/home/openclaw/dev/chat-template-debugger/`) that bypasses all middleware and calls `llm.generate()` directly.

**The #1 priority for this LoRA run** is to make the model emit tool-call tokens natively. Specifically:

1. When the user asks the model to use a tool, the model should emit `startPos` → JSON function call → `endPos` instead of writing `from tools import X` / `X(args)` as Python code
2. This must work for **all** tool patterns — not just structured JSON tools (save_config) but also code-generation tools (write_file) that the model currently code-dumps instead of calling
3. The model should still produce clean text content when NOT invoking a tool — we're adding a capability, not replacing one

### Why this matters

The current "working" save_config path through the vLLM API is not actually the model doing tool calls — the Hermes parser is reconstructing tool calls from the model's text/code output. This is fragile and fails for longer outputs (write_file). Once the model emits native tool-call tokens, both paths work correctly and the parser doesn't need to do salvage work.

## Quick Start

```bash
# Build
docker build -t smollora .

# Run full pipeline (prepare data + train)
docker run --gpus all \
  -v /path/on/host/output:/data/lora-output \
  smollora

# Skip data prep if you already have processed data
docker run --gpus all \
  -e SKIP_PREP=1 \
  -v /path/on/host/processed:/data/processed \
  -v /path/on/host/output:/data/lora-output \
  smollora
```

## Environment Variables

| Var | Default | Description |
|-----|---------|-------------|
| `MODEL` | `HuggingFaceTB/SmolLM3-3B` | Base model (HF repo or local path) |
| `DATA_DIR` | `/data/processed` | Processed data directory |
| `OUTPUT_DIR` | `/data/lora-output` | Training output directory |
| `EPOCHS` | `3` | Training epochs |
| `BATCH_SIZE` | `4` | Per-device batch size |
| `LR` | `2e-4` | Learning rate |
| `LORA_R` | `16` | LoRA rank |
| `MAX_LENGTH` | `4096` | Max sequence length |
| `SKIP_PREP` | `0` | Set to `1` to skip data preparation |

## Datasets

Three datasets combined and converted to SmolLM3's native token format:

1. **interstellarninja/tool-calls-multiturn** — Multi-turn tool calling conversations
2. **NousResearch/Hermes-Function-Calling-V1** — Hermes-format function calling
3. **Salesforce/xLAM-function-calling-60k** — Large-scale function calling (60k samples)

Only conversations containing tool calls are kept. All are normalized to SmolLM3's special tokens:
- Tool calls → `startPos`/`endPos` (token IDs 128002/128016)
- Tool responses → `eni`/`eni_result` (token IDs 128013/128014)

## LoRA Configuration

- **Rank:** 16
- **Alpha:** 32
- **Target modules:** q/k/v/o projections + gate/up/down MLP
- **Dropout:** 0.05
- **Scheduler:** Cosine with 3% warmup
- **Optimizer:** AdamW (fused)
- **Gradient checkpointing:** Enabled

## Output

The trained adapter is saved to `$OUTPUT_DIR/final/`. To use with vLLM:

```bash
# Merge adapter into base model (recommended for vLLM)
python -m peft import PeftModel
# Or pass the adapter path directly with --enable-lora
```

## SSH Deployment

```bash
# On GPU box, after SSH-ing in:
docker run --gpus all -v ~/smol-data:/data smollora

# Or with local model cache:
docker run --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -v ~/smol-data:/data \
  smollora
```