A LoRA adapter for ibm-granite/granite-4.0-350m that turns a natural-language request plus a target JSON Schema into a single schema-conforming JSON object.

This is one of three adapters built to share a single Granite Switch checkpoint (alongside CTI technique-mapping and GenAI attack-vector classification), so its LoRA shape is identical across all three: rank 16, alpha 32, on the fused q/k/v/o attention projections and the input_linear / output_linear MLP projections.

Evaluation

Held-out eval (n = 2000), greedy decoding:

Metric	Score
Key-F1 (field-name agreement vs. gold)	98.4
JSON parse rate	96.8
Schema conformance	94.2
Malformed / empty generations	0

Key-F1 is the headline metric; schema conformance is reported as a diagnostic only (some synthesized gold objects do not validate against their own schema, so it reads low relative to the true generation quality).

Intended use

Constrained JSON generation where the caller supplies the schema in-context. Not a general chat model — use the base Granite model for open-ended conversation.

Quick start

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "ibm-granite/granite-4.0-350m"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="cuda")
model = PeftModel.from_pretrained(model, "barha/granite-text-to-json-350m-lora")

messages = [{"role": "user", "content": "<your request + JSON schema here>"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")
out = model.generate(inputs, max_new_tokens=1024)
print(tok.decode(out[0, inputs.shape[1]:], skip_special_tokens=True))

Training

Base: ibm-granite/granite-4.0-350m
Method: supervised fine-tuning (TRL), LoRA r=16 / α=32, dropout 0.05
Target modules: q_proj, k_proj, v_proj, o_proj, input_linear, output_linear
Data: ChristianAzinn/json-training

License: Apache-2.0