28 days ago
-
index: add g3s (v7 text-only corpus)
-
g3s: final checkpoint iter 18019 — v7 text-only corpus (1.15M rows, 66.5% with reasoning), single variable vs g3r
30 days ago
-
index: add g3q (4 epochs) and g3r (v6 text-only, no tool calls)
-
g3r: final checkpoint iter 14325 — v6 text-only corpus, zero tool-call supervision
-
g3q: final checkpoint iter 9161 — num_epochs 2->4 on v5 (single variable vs g3p)
-
index: add g3p; note dataset_size cannot identify the corpus
-
g3p: final checkpoint iter 4581 — v5 corpus at max_seq_length 65536 (two variables vs g3n)
-
index: g3c final is 6413, not the pre-OOM 6001
-
g3c: TRUE final checkpoint iter 6413 (global_step 6412) — supersedes the pre-OOM 6001 snapshot; the resume wrote to .61/.62, not the original .101/.31
-
index: add g3o; record the rope decomposition
-
g3o: final checkpoint iter 4455 — rope_theta 1e6->5e4 in isolation; theta measured inert (acc +0.00% vs g3n)
-
index: add g3n (default rope vs g3m's yarn)
-
g3n: final checkpoint iter 4455 — g3m with default rope (theta 1e6) instead of yarn
-
index: add g3m (max_seq_length 40000)
-
g3m: final checkpoint iter 4455 — g3f at max_seq_length 40000 (drops 8000 long rows)
-
index: add g3l (v3h6 draft, v34 corpus, 2 epochs)
-
g3l: final checkpoint iter 7527 — v3h6 draft, v34 corpus, 2 epochs (schedule derived, not pinned)
-
g3k: v3h6r finetuned on v4_train, 2 epochs (2946 steps), acc 0.629
-
v3h6r: 4-epoch ablation off v3h4, final iter_0009161 (loss 1.491, acc 0.672)
-
index: add g3j (warm-start from g3e)
-
g3j: final checkpoint iter 4581 — warm-start from g3e iter 4581, g3f architecture (block 3)
-
index: g3g final iter 4581; add the 8-shard requirement
-
g3g: README for final iter 4581; now step-aligned with g3e/g3f
-
g3g: ckpt_meta.json for iter 4581
-
g3g: FINAL_ITERATION for iter 4581
-
g3g: FINAL checkpoint iter 4581 (training completed; supersedes the 3751/4001 in-progress snapshots)
-
g3h: complete model/ with all 8 distcp shards (ranks 4-7 were on the second trainer node)
-
g3g: refresh FINAL_ITERATION
-
g3g: complete model/ with all 8 distcp shards (ranks 4-7 were on the second trainer node)
-
g3d: complete model/ with all 8 distcp shards (ranks 4-7 were on the second trainer node)
-
g3c: complete model/ with all 8 distcp shards (ranks 4-7 were on the second trainer node)
-
g3g: add draft config (late taps [46,57,59])
-
g3h: final checkpoint + run config
-
g3g: final checkpoint + run config
-
g3f: final checkpoint + run config
-
g3e: final checkpoint + run config
-
g3d: final checkpoint + run config
-
g3c: final checkpoint + run config
-
index: the six GB300 g3 arms and their variables
-
initial commit