README.md

---
license: apache-2.0
pipeline_tag: text-generation
tags:
  - lfm2
  - data-use
  - provenance
  - usage
  - impact
  - sft
  - lora
---

# 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=5464. Exact string match of each emitted attribute against the gold label.

| attribute | tp | fp | fn | precision | recall | f0.5 | f1 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| producer | 1081 | 374 | 368 | 0.7430 | 0.7460 | 0.7436 | 0.7445 |
| year | 1002 | 230 | 348 | 0.8133 | 0.7422 | 0.7980 | 0.7761 |
| geography | 1620 | 446 | 407 | 0.7841 | 0.7992 | 0.7871 | 0.7916 |
| acronym | 1270 | 246 | 146 | 0.8377 | 0.8969 | 0.8489 | 0.8663 |
| **overall** | 4973 | 1296 | 1269 | 0.7933 | 0.7967 | 0.7940 | 0.7950 |

Usage/impact macro-F1 (per head):
- data_type: 0.6005
- usage_action: 0.5624
- impact_label: 0.4933
- usage_summary: mean_sim=0.6581 grounded_rate=0.9165

Verbatim rate (emitted values that are substrings of the context): 6249/6269 = 0.9968

## synthetic holdout

Holdout n=1564. Exact string match of each emitted attribute against the gold label.

| attribute | tp | fp | fn | precision | recall | f0.5 | f1 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| producer | 126 | 23 | 26 | 0.8456 | 0.8289 | 0.8422 | 0.8372 |
| year | 123 | 24 | 21 | 0.8367 | 0.8542 | 0.8402 | 0.8454 |
| geography | 163 | 38 | 36 | 0.8109 | 0.8191 | 0.8126 | 0.8150 |
| acronym | 78 | 17 | 13 | 0.8211 | 0.8571 | 0.8280 | 0.8387 |
| **overall** | 490 | 102 | 96 | 0.8277 | 0.8362 | 0.8294 | 0.8319 |

Usage/impact macro-F1 (per head):
- data_type: 0.8559
- usage_action: 0.7706
- impact_label: 0.7985
- usage_summary: mean_sim=0.5276 grounded_rate=0.5898

Verbatim rate (emitted values that are substrings of the context): 592/592 = 1.0000