lfm2.5-350M-datause-multitask
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: 3
- learning rate: 0.0002
- LoRA: r=16 alpha=32 dropout=0.05
- completion-only masking (loss on assistant JSON turn)
Evaluation (holdout, n=5464)
Exact string match of each emitted attribute against the gold label.
| attribute | tp | fp | fn | precision | recall | f0.5 | f1 |
|---|---|---|---|---|---|---|---|
| producer | 1106 | 354 | 343 | 0.7575 | 0.7633 | 0.7587 | 0.7604 |
| year | 1070 | 253 | 280 | 0.8088 | 0.7926 | 0.8055 | 0.8006 |
| geography | 1668 | 425 | 359 | 0.7969 | 0.8229 | 0.8020 | 0.8097 |
| acronym | 1243 | 214 | 173 | 0.8531 | 0.8778 | 0.8580 | 0.8653 |
| overall | 5087 | 1246 | 1155 | 0.8033 | 0.8150 | 0.8056 | 0.8091 |
Usage/impact macro-F1 (per head):
- data_type: 0.6239
- usage_action: 0.5271
- impact_label: 0.4107
Verbatim rate (emitted values that are substrings of the context): 6315/6333 = 0.9972