micic-mihajlo/gemma-4-12b-it-social-post-lora
LoRA adapter for general-purpose social media post drafting.
Base
- Base model:
google/gemma-4-12B-it
- Adapter:
micic-mihajlo/gemma-4-12b-it-social-post-lora
- Training method:
QLoRA + FSDP + MMD/JMQ
Data
- Train examples:
113
- Validation examples:
15
- Generated eval posts:
10
Objective
The adapter is optimized for brief-to-post generation with supervised cross-entropy plus distribution-alignment regularizers over token-logit shape.
Evaluation
| Metric | Value |
|---|
avg_sentence_length | 17.3167 |
avg_word_length | 4.6399 |
char_count | 477.7000 |
cta_marker_count | 0.2000 |
hashtag_count | 0.0000 |
mention_count | 0.0000 |
punctuation_rate | 0.0159 |
sentence_count | 4.8000 |
type_token_ratio | 0.8516 |
url_count | 0.0000 |
word_count | 82.9000 |
Base Comparison
- Style distance improvement:
290.6468
- Adapter-target distance:
190.4017
- Base-target distance:
481.0485
Generation Quality Gate
- Status:
pass
- Failed generations:
0
- Total generations:
10
Limitations
- This adapter should not be used to imitate a private person, brand, or protected identity.
- Quality depends on the human-reference corpus and held-out brief coverage.
- Distribution metrics are diagnostic signals, not a substitute for human review.
