barha/granite-switch-4.0-350m-cti

Granite Switch 4.0 350M — CTI Technique Mapping

A **Granite Switch** model: the `ibm-granite/granite-4.0-350m` base with a single LoRA adapter — `cti_technique_mapping` — embedded into one switchable checkpoint and fired by a control token.

It is the small-model analogue of `ibm-granite/granite-switch-4.1-3b-preview`: same Granite Switch composition machinery (control tokens, KV-hiding, chat-template integration), but built on the 350M Granite 4.0 base and carrying one CTI adapter instead of the preview's granitelib adapter library.

What the adapter does

cti_technique_mapping maps a piece of cyber threat intelligence (CTI) text — a sentence or short passage describing adversary behavior — to the single best-matching MITRE ATT&CK technique ID (e.g. T1059, T1566.001).

The adapter's I/O contract (io_configs/cti_technique_mapping/io.yaml) constrains the output to a single technique-ID string matching ^T[0-9]{4}(\.[0-9]{3})?$, greedy decoding, max_completion_tokens: 16.

The underlying LoRA scored 96.67% exact-match (290/300) on the held-out CTI validation set.

Composition summary

Usage

The control token activates the adapter. With the Granite Switch HF backend:

from granite_switch.hf import GraniteSwitchForCausalLM
from transformers import AutoTokenizer

model_id = "barha/granite-switch-4.0-350m-cti"
tok = AutoTokenizer.from_pretrained(model_id)
model = GraniteSwitchForCausalLM.from_pretrained(model_id, device_map="auto")

cti = "The actor used PowerShell to download and execute a payload from a remote server."
messages = [{"role": "user", "content": cti}]
# The chat template inserts <|cti_technique_mapping|> to fire the adapter.
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=16, do_sample=False)
print(tok.decode(out[0, inputs.shape[1]:], skip_special_tokens=True))
# -> e.g. "T1059.001"

For fast inference, deploy with vLLM (see the granite-switch docs).

How it was built

Composed with granite_switch.composer.compose_granite_switch:

python -m granite_switch.composer.compose_granite_switch \
  --base-model ibm-granite/granite-4.0-350m \
  --adapters <path-to>/cti_technique_mapping \
  --technology lora \
  --output <out>

Note on `granitemoehybrid`: all Granite 4.0 / Nano models are granitemoehybrid configs (with num_local_experts=0), whose MLP leaves are the fused input_linear / output_linear rather than dense gate/up/down_proj. The CTI LoRA was therefore trained targeting input_linear / output_linear (plus attention q/k/v/o_proj) so it composes cleanly — a scalar/all-linear target would produce phantom gate/up/down_proj weights the composer rejects.

Files

• —model.safetensors — composed base + embedded LoRA weights
• —config.json — model_type: granite_switch
• —adapter_index.json — adapter → control-token mapping
• —io_configs/cti_technique_mapping/io.yaml — adapter I/O contract (output schema, decoding params)
• —chat_template.jinja — control-token-aware chat template
• —compose_report.json, BUILD.md — full composition provenance

License

Apache-2.0.