uef-dm128-sft — DM-128 mixed instruction SFT of the two P2@60k d28 checkpoints (kempner)

Weights-only release artifacts (artifact_type: "eval"). Optimizer moments and rng_state are stripped by scripts/export_ckpt_weights.py, so these load for evaluation but are not resumable. Dtypes are preserved verbatim (fp32 in, fp32 out); no cast, no re-pack. Each checkpoint has a .sha256 next to it.

Two identities, one folder each, same recipe and pool; only the frozen image representation and the P2 parent differ:

folder	image representation	init (P2 pretrain @ 60k)	run name
siglip-d28/	google/siglip2-so400m-patch14-224 (1152-d, t_shift 8.485)	Ced-Collab/uef-scaling-a2/camb-p2-siglip-d28/checkpoint_060000.pt	uef_dm128_15k_p2d28_60k_siglip_d28_kempner
dino-d28/	facebook/webssl-dino300m-full2b-224 (1024-d, image_codec: webssl_dino, t_shift 8.0)	Ced-Collab/uef-camb-p2-dino-d28/checkpoint_060000.pt	uef_dm128_15k_p2d28_60k_dino_d28_kempner

Recipe (docs/KEMPNER_DM128_SFT_NOTE.md, configs in each configs/): d28 (2.12B), text T5 (google/flan-t5-small states, txt_len 128), global batch 1024 = micro 8 x grad-accum 16 x world 8 (2 nodes x 4 H200), lr 1e-4, warmup 250, 15,000 steps, seed 4242, bf16 autocast / fp32 master, dual EMA. Per-row answer window (qa_window sidecar: 48 concise captions, 101 detailed captions and long-form answers, global 12 otherwise).

Pool = 20,589,708 rows over 18 manifests (15k x 1024 = 15.36M samples = 0.75 epoch): T2I captions 7,143,157 (blip3o-60k x75, DALL-E 3 x52, ShareGPT-4o x52) · short QA 7,515,498 (VQAv2 train x7, Cambrian v2 gqa/coco/ocr_vqa x2, vg x2, vizwiz x2) · MCQ 486,700 (Cambrian MCQ x10, ScienceQA train x10) · prompted concise captions 2,077,606 · prompted detailed captions 1,616,516 · long-form answers 1,718,221 · coco_valrows 32,010. Site equivalence with the home packs: exact row counts, tar counts and first keys (DM128_COUNTS_OK).

Contents

file	step	bytes	sha256	val vqav2	val long
siglip-d28/checkpoint_005000.pt	5000	25,457,507,318	23ef76c54294af51725c7035838c9d5807c95104442b6d11036ee38b060bbed5	0.1182	0.4928
siglip-d28/checkpoint_015000.pt	15000	25,457,507,190	78121bea64dae50e4fcb75a49cb315c91e9aac898998f2f656100e10cf8ccda8	0.1156	0.5057
dino-d28/checkpoint_005000.pt	5000	25,452,000,758	fedcf4180e61b7ca5682134a2376bc7bc51a8819f3e0f1238203be67449a2b1e	0.1277	0.3030
dino-d28/checkpoint_015000.pt	15000	25,452,000,630	a3b83632d56759c8719bad98021972c694a4e654fc98cace1cc0671ab97b27fc	0.1268	0.3146

checkpoint_005000.pt is the 5k tripwire milestone; checkpoint_015000.pt is the end of the schedule. Validation columns are the in-training monitor only (val/vqav2/loss on the held-out VQAv2-train tars 43-44, val/long/loss on the blip3o long-caption val shards), 4 fixed batches, every 500 steps. They are in each identity's own latent space and are not comparable across the two folders. best_val (not released) was step 13500 for SigLIP2 (vqav2 0.1154) and step 12500 for DINO (vqav2 0.1249). All readouts (linear head, benches, GenEval / DPG, prompted-caption CIDEr) are run elsewhere, not here.

Each folder also carries configs/ (the run config and its resume twin), metrics/ (per-segment metrics.jsonl + launch manifest.yaml) and precision_contract.json.

Segment -> run id -> job id

Each identity is two training segments. Steps 0 -> ~10,150 ran from the P2 parent (init_from, weights only). Steps 10,000 -> 15,000 are a full-state resume from the run's own checkpoint_010000.pt (model, both EMAs, optimizer moments, best_val) through the resume-twin config (..._resume.yml = the run config with init_from removed and auto_resume: true; config and dataset identity hashes equal to the original's). At the segment boundary the WebDataset stream is reseeded to a fresh global permutation.

identity	segment	run id	slurm job	steps	exit
SigLIP2	1	20261003-144439-dm128-siglip	50223968	50 -> 10150	1 (unreadable row, see below)
SigLIP2	2	20261004-144001-dm128-siglip-seg10k	50470446	10050 -> 15000	0
DINO	1	20261003-155217-dm128-dino-r2	50251168	50 -> 10150	cancelled by the stall watchdog (same row)
DINO	2	20261004-162050-dm128-dino-seg10k	50470464	10050 -> 15000	0

Not in the lineage: 20261003-144514-dm128-dino (job 50223970), the first DINO attempt, was cancelled at step 73 because one GPU of its allocation ran a degraded PCIe link and stalled all ranks; it wrote no checkpoint, and the DINO run was restarted from the P2 parent on other nodes. Its one-row metrics.jsonl and manifest are included for completeness. Two resume attempts on 2026-10-04 (jobs 50223969, 50251170) failed at config validation before training anything.

Known notes

• Unreadable rows, repaired between the segments. Both first segments died at step ~10,150 on the same training row: two long-form ALLaVA tar members were named allava_2890691_00.jpg?1 and allava_2897870_00.jpg?1 (the source image path carries a ?1 suffix that the packer kept as the extension), which the loader does not accept as an image. A scan of all 3,620 base tars of the pool found only these two. For the second segments the two members were renamed to .jpg (same bytes, same order, manifests unchanged). Neither row was ever trained on in the first segments.
• Data order. Because of the reseed at step 10,000, the 10k -> 15k samples are not the ones an uninterrupted run would have drawn.
• Pool caveats inherited and added. The P2 parents were trained on the FULL-COCO pool, which includes COCO val rows (coco_valrows, 32,010); this SFT pool includes the same 32,010 rows and VQAv2 train. Every downstream number from this lineage carries that footnote.
• cambrian_vizwiz_x2 holds 40,258 rows at this site against the config's declared 40,256 (one extra row in the site's vizwiz pack, known since P1).