Odia tutor E07: the training checkpoints
Every checkpoint of experiment E07's two fine-tuning runs, kept so any evaluation in the E07 report can be recomputed. The final models, the comparison and the model card are in fastpixels/odia-tutor-e07. Private, research use only.
| Folder | What | Size each |
|---|---|---|
sarvam1-sft/checkpoint-1500, -3000, -4272 |
Sarvam-1 trainer checkpoints: fp32 weights, tokenizer, trainer_state.json (the full log history) |
10.1 GB |
gemma4-e4b-sft/checkpoint-1500, -3000, -4272 |
Gemma-4-E4B-it trainer checkpoints: trained weights in fp32, frozen parts (per-layer embeddings, vision, audio) in bf16 | about 27 GB |
gemma4-e4b-sft/step-800-bf16, -1600-bf16, -2300-bf16 |
Gemma snapshots at the steps matched to Sarvam-1's checkpoints (same tokens seen, about the same hours): bf16 weights and tokenizer | 17.3 GB |
Only the newest checkpoint of each run keeps its resume state (optimizer.pt with AdamW's moments, scheduler.pt,
rng_state.pth): it is uploaded with every checkpoint while training runs, as a failsafe, and removed from the
older one once a newer checkpoint has its own. At the end, checkpoint-4272 keeps it (Sarvam-1 20.2 GB, Gemma
about 37 GB), so either run can be continued exactly. Each run is 4,272 steps: two passes over 136,664 rows in
batches of 64.
Use
Download one folder, never the whole repo (a few hundred GB):
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer
repo, sub = "fastpixels/odia-tutor-e07-checkpoints", "sarvam1-sft/checkpoint-3000"
path = snapshot_download(repo, allow_patterns=f"{sub}/*", ignore_patterns=["*optimizer.pt"])
model = AutoModelForCausalLM.from_pretrained(f"{path}/{sub}", dtype="bfloat16")
tok = AutoTokenizer.from_pretrained(f"{path}/{sub}")
To continue training from the last checkpoint, download checkpoint-4272 with its optimizer.pt into the run's
output folder and resume (odia-train ... --resume, or Trainer.train(resume_from_checkpoint=...)). The cosine
schedule ended at step 4,272, so going further needs a new schedule.
Licence
sarvam1-sft/: the Sarvam AI Research License (non-commercial), inherited from Sarvam-1. gemma4-e4b-sft/: base
model Apache 2.0; research use only because of the fine-tuning data (SciQ is CC BY-NC 3.0).