stage-2 controlled-phase CONTROL arm (stock trunk, 0-2.5k) - anchors the published paired experiment - no optimizer

flow rig ft BEST (12.005 @9.2k) - no optimizer

flow rig ft 10k extension final (best 12.005 @9.2k, final 12.098; 4k phase best was 12.229; AR ft floor 12.89) - no optimizer

flow rig ft 4k (stage-2 flow-80k base): best 12.229 @3600 vs AR ft floor 12.89; frozen trunk resists memorization (train 5.2 vs 1.0) - no optimizer

stage-2 flow on AR trunk, 80k extension final (in-run probe 6.574; beats the h1536 lineage's 100k endpoint at 80% of steps) - WITH optimizer

stage-2 flow on AR trunk, 40k final (eval 7.104; ties the h1536 lineage's 100k at 40% of steps) - WITH optimizer

rig ft best-by-eval: step 800 (eval_chunk_mae 12.892), no optimizer

rig ft (rejudged opus-5 labels): rcond-100k + 4k steps, no optimizer

prune stale optimizer.pt files (older checkpoints will not be resumed; rcond 100k optimizer is the only live one)

bijou_arb_rcond_100k_ddp4 step 100000 (with optimizer)

Upload folder using huggingface_hub

Upload bijou_ft_rig_from_adarms100k_ddp2/step_002500/expert.safetensors with huggingface_hub

Upload bijou_ft_rig_from_adarms100k_ddp2/step_002500/bijou_config.json with huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

unftext r2 continuation, step 25000: unfrozen-text expert + adapted backbone (no optimizer)

unftext15k r2: unfrozen-text-trunk expert + adapted backbone at step 15000 (no optimizer)

FAST action tokenizer v1: DCT+BPE, fit on 1040 datasets / 4.9M chunks (community v1/v2/v3 + rig), vocab 1024

bijou_ft_marius_4k_init45k/step_004000: expert weights + optimizer state + config

bijou_community_v1v2v3_cont45k_ddp4/step_045000: expert weights + optimizer state + config

bijou_ft_marius_2k: drop optimizer states (easy to repro, ~26GB)

bijou_ft_marius_2k/step_002000: expert weights + optimizer state + config

bijou_ft_marius_2k/step_001750: expert weights + optimizer state + config

bijou_ft_marius_2k/step_001500: expert weights + optimizer state + config

bijou_ft_marius_2k/step_001250: expert weights + optimizer state + config

bijou_ft_marius_2k/step_001000: expert weights + optimizer state + config

bijou_ft_marius_2k/step_000750: expert weights + optimizer state + config

bijou_ft_marius_2k/step_000500: expert weights + optimizer state + config

bijou_ft_marius_2k/step_000250: expert weights + optimizer state + config

bijou_community_v1v2v3_20k_ddp4/step_040000: expert weights + optimizer state + config

bijou_community_v1v2v3_20k_ddp4/step_015000: expert weights + optimizer state + config

bijou_community_v1v2_20k_ddp4/step_010000: expert weights + optimizer state + config

bijou_community_v2_20k_pdnorm_h1024_r2/step_002500: expert weights + optimizer state + config

model card