LoRA adapter for ethical copywriting for the Brazilian third sector
(ONGs, OSCIPs, institutes, and social foundations), fine-tuned on
Mixtral-8x7B-Instruct-v0.1 via Adaption's
AutoScientist platform.
Covers 12 distinct communication formats across fundraising, ESG proposals,
impact reports, campaigns, social media, grant applications, beneficiary
communication, volunteer mobilization, transparency notes, advocacy,
and crisis communication — all in PT-BR.
Evaluation results
Training Winrates
The adapted model outperforms the base model with 69% win rate vs 31%
on held-out copywriting evaluation — a +122% relative improvement
over the base model.
Model
Win Rate
Base (Mixtral-8x7B-Instruct-v0.1)
31%
Adapted (causabr-impactcopy)
69%
Train/Eval Metrics
Metric
Value
Initial train loss
~2.00
Final validation loss
~0.93
Loss reduction
~54%
Peak learning rate
1.0e-6
Training steps
204
LR scheduler
cosine (warmup)
Gradient norm
spike → stable
Train loss (cyan) converged steadily over 204 steps.
Validation loss (black dots) tracked closely, confirming generalization
without overfitting. Learning rate followed cosine schedule with warmup,
peaking around step 30 then decaying to near-zero. Gradient norm stabilized
after initial spike, indicating stable optimization throughout.
Experimental research artifact submitted to the AutoScientist Challenge
2026 (Marketing category). Generated copy requires human review before
publication. The dignity gate enforces ethical communication standards
but does not substitute for editorial judgment by communications
professionals.
