lilly-fashion-muse-absa

DistilBERT ABSA (aspect consensus)

Lilly · Fashion Muse — aspect side. Aspect-level sentiment for quality / price / fit (shipping intentionally excluded) on fashion e-commerce reviews. Input format: sentence + aspect pair.

Provenance

Trained on aspect-sentiment pairs built by a doc+sentence consensus filter (both the doc-level model and a sentence-level model must agree on the sentence's aspect polarity) over weak labels derived from star ratings.

Measured performance

• Human gold subset (n = 91 pairs): acc .45 / macroF1 .45 — small sample, and the same ~0.15 annotation-kappa ceiling applies. Treat as directional.
• Consensus filtering trades recall for label cleanliness by design.

Honest limitations

• Weak-label provenance (star-derived); consensus reduces but does not eliminate label noise.
• Only quality / price / fit aspects are supported.
• Max sequence length 96. Labels: 0=negative, 1=neutral, 2=positive.

Artifacts & loading

Same fp16-shard format as the doc model (merge shards, strict load_state_dict, .float().eval()). Verified 100.0% round-trip agreement.