ladder-fmt05 r233: the fix-ladder's working RL recipe

ep1-ppr32: full-epoch stock-algorithm RL — collapse by r140; rollouts + metrics

ppr32-scope: 60-round single-GPU throughput scope, 512 rollouts/round @ 2.1s

align10: 10-round aligned chat_rl_fast smoke off the tool-free SFT ckpt (rollouts + metrics)

Tool-free SFT of d24_decoderstack step 5568 (462 steps, val bpb 0.2732, ChatCORE 0.2527); model + meta + w1 optim shard

Model card: correct the recalibration — leaderboard is best-fit only, so the like-for-like gap is +0.019; both checkpoints now measured under both packings

Model card: recalibrate val bpb under upstream packing; document the token_bytes regeneration

tokenizer: keep the pre-fix token_bytes so published bpb numbers stay reproducible

tokenizer: regenerate token_bytes.pt with upstream 2ce972a raw-bytes fix (191/32768 ids corrected)

Model card: step 5568 is now a ready-made nanochat checkpoint (+8-rank optimizer shards)

meta_005568.json: add user_config so chat_sft inherits the run's LRs

meta_001950.json: add user_config so chat_sft inherits the run's LRs

Add nanochat-format d24 checkpoint (step 5568) + world_size=8 optimizer shards

Flag that val bpb is not comparable across the batched/varlen pipelines

Converter now emits nanochat optimizer shards for an SFT warm-start

Correct optimizer-state notes: SFT needs only the weights; W_O axis is a no-op at d24

Add model card, meta.json, logs, tokenizer, training source, and nanochat converter

Upload folder using huggingface_hub

initial commit