Qwen3.5-4B-en-es-fa-hi-ron

Qwen3.5-4B day-by-day multilingual SFT (en -> es -> fa -> hi -> ro)

Qwen3.5-4B fine-tuned sequentially on 5 languages for a continual-learning "day in the life" experiment. Each language trained for one "day" (3 epochs, 1000 train rows), starting from the previous day's checkpoint. This checkpoint is the end of day 5 (Romanian) - the model has seen all 5 languages in order.

Task: Fill [MASK] tokens in an "I feel ..." sentence with the corresponding affective state expression(s).

Results - set_acc@1 on held-out test sets (1000 rows/lang)

                test_en   test_es   test_fa   test_hi   test_ro
day1 (en)       0.187*    0.149     0.114     0.117     0.121
day2 (->es)     0.192     0.267*    0.112     0.088     0.090
day3 (->fa)     0.166     0.231     0.445*    0.068     0.111
day4 (->hi)     0.157     0.234     0.315     0.571*    0.174
day5 (->ro)     0.144     0.167     0.358     0.533     0.443*

* = just-trained language. This checkpoint corresponds to the final row.

Training details

• Base: Qwen/Qwen3.5-4B
• Order: en -> es -> fa -> hi -> ro (one language per day)
• Per day: 3 epochs, effective bs 4 (per_device=1, grad_accum=4)
• LR 1e-5, cosine, warmup 0.03, bf16, sdpa attention, gradient checkpointing
• Single A100-40GB SXM

Code and full run log

https://github.com/Continual-Learning-Emotion-Group/Romanian_ASI/tree/day-by-day See pipeline/train/RUN_LOG_DAY_BY_DAY.md.