LoRA SFT of LiquidAI/LFM2.5-350M for data-mention provenance attributes (producer / year / geography / acronym) and usage/impact classification (data_type / usage_action / impact_label / usage_summary).

Training

base model: LiquidAI/LFM2.5-350M
dataset: rafmacalaba/data-use-sft
epochs: 5
learning rate: 0.0002
LoRA: r=16 alpha=32 dropout=0.05
completion-only masking (loss on assistant JSON turn)

real holdout

Holdout n=14776. Exact string match of each emitted attribute against the gold label.

attribute	tp	fp	fn	precision	recall	f0.5	f1
producer	1909	968	914	0.6635	0.6762	0.6660	0.6698
year	2264	520	631	0.8132	0.7820	0.8068	0.7973
geography	2632	984	968	0.7279	0.7311	0.7285	0.7295
acronym	1973	456	449	0.8123	0.8146	0.8127	0.8134
overall	8778	2928	2962	0.7499	0.7477	0.7494	0.7488

Usage/impact macro-F1 (per head):

data_type: 0.7005
usage_action: 0.6554
impact_label: 0.5842
usage_summary: mean_sim=0.5522 grounded_rate=0.6513

Verbatim rate (emitted values that are substrings of the context): 11659/11706 = 0.9960

synthetic holdout

Holdout n=2305. Exact string match of each emitted attribute against the gold label.

attribute	tp	fp	fn	precision	recall	f0.5	f1
producer	231	39	56	0.8556	0.8049	0.8449	0.8294
year	265	50	40	0.8413	0.8689	0.8466	0.8548
geography	344	63	53	0.8452	0.8665	0.8494	0.8557
acronym	145	84	36	0.6332	0.8011	0.6609	0.7073
overall	985	236	185	0.8067	0.8419	0.8135	0.8239

Usage/impact macro-F1 (per head):

data_type: 0.7818
usage_action: 0.6489
impact_label: 0.5147
usage_summary: mean_sim=0.5752 grounded_rate=0.6768

Verbatim rate (emitted values that are substrings of the context): 1216/1221 = 0.9959

rafmacalaba/lfm2.5-350M-multitask-datause

lfm2.5-350M-multitask-datause

Training

real holdout

synthetic holdout

rafmacalaba/
lfm2.5-350M-multitask-datause