---
base_model: ibm-granite/granite-4.0-350m
library_name: transformers
pipeline_tag: text-generation
tags:
- granite
- granite-switch
- lora
- adapters
- mixture-of-adapters
license: apache-2.0
---

# Granite Switch 4.0 350M — 3-Adapter Demo

A single [Granite Switch](https://github.com/generative-computing/granite-switch)
checkpoint built on [`ibm-granite/granite-4.0-350m`](https://huggingface.co/ibm-granite/granite-4.0-350m)
with **three task LoRA adapters embedded in one model**. Each adapter is activated
by a control token, so one deployed checkpoint serves three different tasks — no
adapter swapping, no separate model loads.

This is the multi-adapter companion to the single-adapter
[`barha/granite-switch-4.0-350m-cti`](https://huggingface.co/barha/granite-switch-4.0-350m-cti).

## Embedded adapters

| # | Adapter | Control token | Task | Source LoRA |
|---|---------|---------------|------|-------------|
| 1 | `cti-technique-mapping` | `<\|cti-technique-mapping\|>` | Map a CTI description → MITRE ATT&CK technique ID | [barha/granite-cti-technique-mapping-350m-lora](https://huggingface.co/barha/granite-cti-technique-mapping-350m-lora) |
| 2 | `text-to-json` | `<\|text-to-json\|>` | Natural language + schema → schema-conforming JSON | [barha/granite-text-to-json-350m-lora](https://huggingface.co/barha/granite-text-to-json-350m-lora) |
| 3 | `genai-attack-vector` | `<\|genai-attack-vector\|>` | Classify a GenAI security incident into 1 of 14 attack-vector classes | [barha/granite-genai-attack-vector-350m-lora](https://huggingface.co/barha/granite-genai-attack-vector-350m-lora) |

All three adapters share an identical LoRA shape (rank 16, alpha 32, on the fused
q/k/v/o attention projections and the `input_linear` / `output_linear` MLP
projections), which is what lets them stack cleanly into one switch checkpoint.

Control token IDs: `100352`, `100353`, `100354` (3 new tokens; vocab 100355).

## Per-adapter evaluation (on granite-4.0-350m)

| Adapter | Metric | Score | n |
|---------|--------|-------|---|
| text-to-json | Key-F1 (headline) | **98.4** | 2000 |
| genai-attack-vector | Accuracy | **74.8%** (166/222) | 222 |
| cti-technique-mapping | — | see source repo | — |

These are the standalone scores of each source LoRA; embedding them in the switch
does not change adapter weights.

## How adapter selection works

Granite Switch routes per request via the control token: place the adapter's
control token in the prompt (the chat template handles placement) and the switch
activates that adapter for the turn. With no control token, the base model runs
unmodified.

## Usage

Compose / inference follow the standard Granite Switch flow — see the
[granite-switch repo](https://github.com/generative-computing/granite-switch) and
its tutorials. The checkpoint loads as a `GraniteSwitchForCausalLM`; `adapter_index.json`
lists the adapter → control-token mapping and `io_configs/<adapter>/io.yaml` carries
each adapter's I/O contract.

## Build

Composed with `granite_switch.composer.compose_granite_switch`:

```bash
python -m granite_switch.composer.compose_granite_switch \
  --base-model ibm-granite/granite-4.0-350m \
  --technology lora \
  --adapters cti-technique-mapping text-to-json genai-attack-vector \
  --output ./granite-switch-4.0-350m-demo
```

Base params 352M → composed 359M (+2.0%). See `BUILD.md` and `compose_report.json`
in this repo for the full composition report.

License: Apache-2.0