nchapman/granite-4.0-1b-summarization-lora

Granite 4.0-1B Summarization LoRA

An exploratory rank-16 LoRA adapter for English web-page summarization, trained from ibm-granite/granite-4.0-1b. This is a PEFT adapter, not a standalone model; load it with the base model.

Evaluation snapshot

Fresh aligned 250-article CNN/DailyMail comparison using Q4 base inference, a BF16 adapter sidecar, common greedy summary-v3 generation, and the same pinned judge used for all four base/adapter pairs:

Metric	Base	Base + adapter	Difference [95% paired CI]
Factuality (judge, 1–5)	4.304	4.328	+0.024 [-0.112, +0.156]
Coverage (judge, 1–5)	4.052	3.808	-0.244 [-0.324, -0.168]
Concision (judge, 1–5)	3.936	4.116	+0.180 [+0.096, +0.260]
Overall quality (judge, 1–5)	4.036	4.000	-0.036 [-0.140, +0.068]
ROUGE-1	0.3219	0.3454	+0.0235 [+0.0150, +0.0323]
Mean words	154.9	141.9	-12.9 [-24.3, +0.2]
1,024-token cap stops	0	2	+2

The result is mixed: reference overlap and concision increased, coverage decreased, and factuality/overall-quality changes were inconclusive. Cap stops rose from zero to two. It is an exploratory adapter, not a quality upgrade or deployment recommendation. Full dimensions, protocol, intervals, and limitations: docs/results/aligned-adapter-comparison.md in the Lillm Summarizer repository. This shared-prompt result uses a common pinned judge and supersedes the earlier Granite snapshot's mutable judge alias.

Training and data

LoRA rank 16, alpha 16, dropout 0; evaluated artifact granite4-1b-summary-v3-8k-s42-early512/checkpoints/checkpoint-128. The training examples pair source documents with target summaries. The public component came from the HuggingFaceTB/smoltalk2 SFT summarization split, which includes one-line and three-sentence summary styles; the mixed runs also used the author's private web-document/summary pairs.

The adaptation goal is to reinforce a shared summarization response contract—objective plain prose, brief when the source is simple and more detailed when it warrants it—across model sizes and families. This goal is not yet established as a cross-model normalization result: no formal style-variance metric was run, and the two Granite adapters show mixed, model-dependent quality outcomes. Private records and derived training files are not included here. Training used one seed; the adapter is not an exhaustive checkpoint or seed search.

Use

Load the adapter with PEFT and the base Transformers model. The published comparison itself used a Q4_K_XL GGUF base plus a separately exported BF16 LoRA sidecar, so native Transformers generations may differ.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "ibm-granite/granite-4.0-1b"
adapter_id = "nchapman/granite-4.0-1b-summarization-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

Use the base model's tokenizer/chat template. The adapter does not bundle a tokenizer or base-model weights.

Licensing

Adapter weights are released under Apache-2.0. The base model retains its own license and terms; consult Granite 4.0-1B before use.