ViuAI-500M / code

Commit History

Today

  1. Update pretrain v2 shard builder with 10B 5-pillar mix (25% Chat, 10% Hinglish)

    Verified signature
    4851e0f

4 days ago

  1. Upload code/train_dpo.py with huggingface_hub

    Verified signature
    1f9c7a2
  2. Upload code/train_dpo.py with huggingface_hub

    Verified signature
    72966e6
  3. Upload code/train_align.py with huggingface_hub

    Verified signature
    39117e7

5 days ago

  1. Upload code/train_dpo.py with huggingface_hub

    Verified signature
    b3c38d7
  2. Upload code/build_dpo_dataset.py with huggingface_hub

    Verified signature
    342b142
  3. Upload code/train_dpo.py with huggingface_hub

    Verified signature
    1e181f0
  4. Upload code/train_dpo.py with huggingface_hub

    Verified signature
    e551219
  5. Upload code/train_sft.py with huggingface_hub

    Verified signature
    d031ac3
  6. Upload code/train_sft.py with huggingface_hub

    Verified signature
    51d6f04
  7. Upload code/train_sft.py with huggingface_hub

    Verified signature
    a12a68c
  8. Upload code/train_sft.py with huggingface_hub

    Verified signature
    bb3bb0d

6 days ago

  1. Upload code/data_builders/build_500m_sarus_1m_dataset.py with huggingface_hub

    Verified signature
    dda27ab
  2. Upload code/train_sft.py with huggingface_hub

    Verified signature
    6c7fabe
  3. Upload code/data_builders/build_500m_sarus_1m_dataset.py with huggingface_hub

    Verified signature
    0c19fce

11 days ago

  1. Update code/build_pretrain_v2_shards.py for Pretraining V2 Continuation

    Verified signature
    522dd81
  2. Update code/train.py for Pretraining V2 Continuation

    Verified signature
    323f259

  1. Update train_sft.py with Pure English v2 anchors and domain mapping

    Verified signature
    0e2a7ef

  1. Update train_dpo.py for calibrated DPO v2 training

    Verified signature
    36191e7
  2. Security: remove hardcoded tokens from code/build_dpo_dataset.py, use env/arg only

    Verified signature
    fb8f4c4
  3. Security: remove hardcoded tokens from code/train_dpo.py, use env/arg only

    Verified signature
    b481f47
  4. Add SFT auxiliary loss (0.5x) and boost DPO learning rate to 1.5e-5 for 500M model

    Verified signature
    e81c88f
  5. Upload latest DPO v1 alignment script: code/build_dpo_dataset.py

    Verified signature
    6a3ea9c
  6. Upload latest DPO v1 alignment script: code/train_dpo.py

    Verified signature
    f3cbfc2
  7. Upload code/train_sft.py with huggingface_hub

    Verified signature
    e9edafc
  8. Delete code/build_sft_v2_chat_dataset.py with huggingface_hub

    Verified signature
    8a7fbdc
  9. Delete code/build_sarus_sft_dataset_kaggle.py with huggingface_hub

    Verified signature
    322b3c6
  10. Delete code/build_dpo_dataset.py with huggingface_hub

    Verified signature
    78fc1c5
  11. Upload code/config.py with huggingface_hub

    Verified signature
    29474cb
  12. Upload code/train_sft.py with huggingface_hub

    Verified signature
    d73ea55
  13. Update code/tokenize_sft_v1_kaggle.py for SFT v1 21-domain curriculum

    Verified signature
    074f578
  14. Update code/train_sft.py for SFT v1 21-domain curriculum

    Verified signature
    aa6f9f5

  1. Only push final checkpoint to HF to save 10 min upload time

    Verified signature
    89de0ef
  2. Upload both final and best checkpoints to Hugging Face

    Verified signature
    f8754ae
  3. Update code/build_sarus_sft_dataset_kaggle.py with user enhancements

    Verified signature
    78132cf
  4. Update code/build_dpo_dataset.py with user enhancements

    Verified signature
    1cdadee
  5. Update code/train_dpo.py with user enhancements

    Verified signature
    22cefb1
  6. Update code/model.py with user enhancements

    Verified signature
    9a49ef4
  7. Update code/build_sft_v2_chat_dataset.py with user enhancements

    Verified signature
    28a4391
  8. Update code/train_sft.py with user enhancements

    Verified signature
    ffb997d
  9. Upload code/build_sft_v2_chat_dataset.py with huggingface_hub

    Verified signature
    c6e9625
  10. Upload code/build_sft_v2_chat_dataset.py with huggingface_hub

    Verified signature
    3ca8d76
  11. Upload code/train_sft.py with huggingface_hub

    Verified signature
    a0c3ea3
  12. Update code/train_sft.py with chunked cross-entropy and H200 turbo speed optimizations

    Verified signature
    2f86ee7
  13. Update code/model.py with chunked cross-entropy and H200 turbo speed optimizations

    Verified signature
    16e21dc
  14. Update code/model.py with chunked cross-entropy and H200 turbo speed optimizations

    Verified signature
    5d8dea0
  15. Update code/train_sft.py with chunked cross-entropy and H200 turbo speed optimizations

    Verified signature
    e4d110c
  16. Update code/model.py with chunked cross-entropy and H200 turbo speed optimizations

    Verified signature
    06b7264
  17. Enable pure FlashAttention-2 & 8 async dataloader workers

    Verified signature
    90b2f87
  18. Update H200 ultra-tier settings

    Verified signature
    fe219e8