AETHORIA-AI/tr-hash-200m-70b-replay-checkpoints

TR-HASH 200M — 70B/130B Replay Checkpoints (run in progress)

Status: run in progress. This is a raw checkpoint backup repository, not a finished model release. It exists as a safety net against local disk loss — every checkpoint here is a resumable training state (model weights, optimizer state, scheduler state), not a validated or evaluated export.

• Checkpoints are uploaded automatically at token-pack boundaries and on clean/interrupted shutdown by scripts/sync_checkpoints_to_hf.py.
• Folder names follow {tag}_{step} (token_pack_NNN_STEP, final_STEP, interrupted_STEP).
• No checkpoint here has a full validated evaluation. Do not treat any of these as a finished model.
• The architecture config is tracked at model_config.yaml in this repo (not embedded in checkpoint.pt).

Exploratory zero-shot check (not a full evaluation)

One informal check on token_pack_016_66120 (~52B tokens trained so far, ~40% of the planned 130B), via scripts/convert_to_mlx.py + scripts/eval_mlx_zero_shot.py, zero-shot causal-log-likelihood scoring, no chat template:

Benchmark	acc	acc_norm	examples
PIQA	0.6567	0.6589	1,838

This is a single mid-training data point on one benchmark, not a release evaluation — numbers will move as training continues.

The finished, evaluated model will be published separately under AETHORIA-AI once the run completes.

See Complexity Framework for the training code.