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)
real holdout
Holdout n=5471. Exact string match of each emitted attribute against the gold label.
| attribute | tp | fp | fn | precision | recall | f0.5 | f1 |
|---|---|---|---|---|---|---|---|
| producer | 1093 | 393 | 356 | 0.7355 | 0.7543 | 0.7392 | 0.7448 |
| year | 1034 | 278 | 316 | 0.7881 | 0.7659 | 0.7836 | 0.7769 |
| geography | 1612 | 476 | 415 | 0.7720 | 0.7953 | 0.7766 | 0.7835 |
| acronym | 1238 | 217 | 178 | 0.8509 | 0.8743 | 0.8554 | 0.8624 |
| overall | 4977 | 1364 | 1265 | 0.7849 | 0.7973 | 0.7874 | 0.7911 |
Usage/impact macro-F1 (per head):
- data_type: 0.5934
- usage_action: 0.5131
- impact_label: 0.4816
- usage_summary: mean_sim=0.6700 grounded_rate=0.9264
Verbatim rate (emitted values that are substrings of the context): 6317/6341 = 0.9962
synthetic holdout
Holdout n=575. Exact string match of each emitted attribute against the gold label.
| attribute | tp | fp | fn | precision | recall | f0.5 | f1 |
|---|---|---|---|---|---|---|---|
| producer | 123 | 27 | 26 | 0.8200 | 0.8255 | 0.8211 | 0.8227 |
| year | 119 | 20 | 35 | 0.8561 | 0.7727 | 0.8380 | 0.8123 |
| geography | 165 | 36 | 29 | 0.8209 | 0.8505 | 0.8267 | 0.8354 |
| acronym | 62 | 26 | 27 | 0.7045 | 0.6966 | 0.7029 | 0.7006 |
| overall | 469 | 109 | 117 | 0.8114 | 0.8003 | 0.8092 | 0.8058 |
Usage/impact macro-F1 (per head):
- data_type: 0.9478
- usage_action: 0.8890
- impact_label: 0.6645
- usage_summary: mean_sim=0.7160 grounded_rate=0.9827
Verbatim rate (emitted values that are substrings of the context): 575/578 = 0.9948