Continual Learning - Round 1 Results

Summary

Successfully trained v11 model on combined old + new data. Val loss improved from 2.476 to 2.051 in just 50 steps.

Data Mix

• Old data (sft_r3.jsonl): 17,273 records (85.5%)
• New data (reviewed/ready): 2,926 records (14.5%)
• Total: 20,199 records

Training Setup

• Init from: r3b checkpoint (val loss 2.476)
• LR: 1e-5
• Val frac: 0.15
• Batch size: 4
• Max iters: 50 (test run)

Results

Step 0:  train 3.0181 | val 2.9935
Step 10: train 2.1656 | val 2.0826
Step 20: train 2.0381 | val 2.0588
Step 30: train 1.9763 | val 2.0684  <- best
Step 40: train 1.9883 | val 2.0512  <- best
Step 50: train 2.0142 | val 2.0718

Best val loss: 2.0512 (17% improvement over r3b)

Next Steps

1. Run full 500 iterations on r1
2. Test the checkpoint with chat.py
3. Evaluate response quality vs r3b
4. If good, proceed to round 2 with reviewed batch data

Files Created

• v11/mkm_train/prepare_continual_data.py - Data preprocessing script
• v11/data/continual_round1.jsonl - Combined dataset
• v11/models/sft_round2/r1_test/ - Test checkpoint
• v11/notes/continuous-learning.md - Updated documentation