Coffee Capsules Core Diffusion Sweep

This repository packages the coffee_capsules_category1_core_diffusion sweep from multitask_dit_policy.

What Is Included

This sweep compares DDPM vs DDIM, 20 vs 50 inference steps, and RoPE vs no-RoPE.

For each of the 8 variants, this repo publishes two params-only checkpoints: 40000 and 50000.

Each published checkpoint includes model.safetensors, config.json, and ramen_stats.pt. train_state.pt is intentionally omitted to keep the sweep inference-focused.

Main Takeaway

All 8 sweep variants trained cleanly and outperformed the older coffee_capsules_v1 baseline at step 50000, which ended around train/loss ~= 0.011694.

Within this sweep, the only consistent split was RoPE vs non-RoPE: RoPE variants finished at 0.0029481, while non-RoPE variants finished at 0.0030173. That difference is small enough that downstream evaluation quality should be the real decision-maker.

Variant Summary

variant	50k train loss	canonical W&B	note
`baseline`	`0.0030173`	8xhewlqu	Control run.
`ddpm_20`	`0.0030173`	leamkw9q	No clear gain over baseline.
`ddpm_50`	`0.0030173`	ay6nbct0	No clear gain over baseline.
`ddpm_rope`	`0.0029481`	3695rhof	Best-loss group.
`ddim_20`	`0.0030173`	innf1jzx	Stable, but tied with baseline group.
`ddim_50`	`0.0030173`	266664n6	Stable, but tied with baseline group.
`ddim_rope_20`	`0.0029481`	qh7lwdtq	Best-loss group.
`ddim_rope_50`	`0.0029481`	1j441mvr	Best-loss group.

Integrity Verification

All checksums below were generated twice directly from the source checkpoint files on Isambard. The values were only recorded after the two passes matched exactly.

variant	step	model.safetensors	config.json	ramen_stats.pt	manifest_sha256
`baseline`	`40000`	`4f45b07f3a8267fdf2b5fcfc02ce94a655ea5cd6d4da285ca69caf75780dc727`	`1b1c785af55dd820ef70820dbc8955cb3b3425e11173fa936a86571876607ed5`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`50bcb62b44fbc4e92cdce8da40ded1fa1a6d62c074747522dd0a85ac3faa9e40`
`baseline`	`50000`	`1e98c59755b247dd06ef2e56086fc2dd9db531a722b2535a760c2654906c6d38`	`1b1c785af55dd820ef70820dbc8955cb3b3425e11173fa936a86571876607ed5`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`76c9085eb40d95d6fc181d90c81102efaa6743232165ec90c3ba522f805af1a8`
`ddpm_20`	`40000`	`4f45b07f3a8267fdf2b5fcfc02ce94a655ea5cd6d4da285ca69caf75780dc727`	`014b2785ed96ec14b923f4b708df15370aef4907af184fa22c69b921d41c3a20`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`6c4210b1af83233b5c433db8314f52604ba8cb10aed282ed590704012d7312af`
`ddpm_20`	`50000`	`1e98c59755b247dd06ef2e56086fc2dd9db531a722b2535a760c2654906c6d38`	`014b2785ed96ec14b923f4b708df15370aef4907af184fa22c69b921d41c3a20`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`7443ab993ed13d417f9ed3a4a5f7bd45c3cc0b3f037be9bdebccd175bc86c656`
`ddpm_50`	`40000`	`4f45b07f3a8267fdf2b5fcfc02ce94a655ea5cd6d4da285ca69caf75780dc727`	`0d963b408b541f9b75b79fbe10a1462c9ecca034125c8b2c178d79b637cf975d`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`5bf388a6e688096253b3c7eabe8679a9551f52bd5bc170bdf4c97521b251dbd1`
`ddpm_50`	`50000`	`1e98c59755b247dd06ef2e56086fc2dd9db531a722b2535a760c2654906c6d38`	`0d963b408b541f9b75b79fbe10a1462c9ecca034125c8b2c178d79b637cf975d`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`734b59f0f69627f3810ddc075af93a3b067059f099fc1674a2d47bdeebfcb6a0`
`ddpm_rope`	`40000`	`d4c50efe6711cf387e3801623bee554366908613346f6c49c978e99a65a9f1aa`	`baf684d0e35f60057cff63db8e91382522af67174985c427d31960f6d7bf2ced`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`9a3f247af34d20ca8a9d832a07fc485f65039ae0ec659614bec819a898c5a45f`
`ddpm_rope`	`50000`	`fcb539e3a2c269cd964756b820697be031ca689fbee8d30906fdffba6ab78dd1`	`baf684d0e35f60057cff63db8e91382522af67174985c427d31960f6d7bf2ced`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`ccd70303182bde4dd2a88dea78f6bea5bcbb166bd73e51ab25fbb381b9d215b8`
`ddim_20`	`40000`	`4f45b07f3a8267fdf2b5fcfc02ce94a655ea5cd6d4da285ca69caf75780dc727`	`8787253b3315853707692775014b8d7f7dbb1298eed132fcd521b176cbbbc9fe`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`cabd203413965463d18b69509f1bdcdfad46f719ec5c790b00c8a5c680824a01`
`ddim_20`	`50000`	`1e98c59755b247dd06ef2e56086fc2dd9db531a722b2535a760c2654906c6d38`	`8787253b3315853707692775014b8d7f7dbb1298eed132fcd521b176cbbbc9fe`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`954cf50d6314f8dcc3fa257eb9ba87013ed2c6e36619115f5d5e5026411331e6`
`ddim_50`	`40000`	`4f45b07f3a8267fdf2b5fcfc02ce94a655ea5cd6d4da285ca69caf75780dc727`	`3d3ebbcac48342d1504f8518e5392862d6fbffbdb9e3d84bcba53dfc7609f94a`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`e144c267dd8b40d0d89667aedd4f76dc296b60c4a6dbcede6fb835b473d63af9`
`ddim_50`	`50000`	`1e98c59755b247dd06ef2e56086fc2dd9db531a722b2535a760c2654906c6d38`	`3d3ebbcac48342d1504f8518e5392862d6fbffbdb9e3d84bcba53dfc7609f94a`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`e667dcd98ee1bc495b454d45df62efee674c2bfd29ec89d02fad711c91db446c`
`ddim_rope_20`	`40000`	`d4c50efe6711cf387e3801623bee554366908613346f6c49c978e99a65a9f1aa`	`ebde4b7349317b24185799fa0ac9b5110a42b2f8ba46de970f352acb8291e095`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`5e7a66d0e3c05e357c675bed4d6a112e31748e3843d047befed22c4a7c220eaf`
`ddim_rope_20`	`50000`	`fcb539e3a2c269cd964756b820697be031ca689fbee8d30906fdffba6ab78dd1`	`ebde4b7349317b24185799fa0ac9b5110a42b2f8ba46de970f352acb8291e095`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`262b9839cb1124e9d39e07cf5762a0d673dc48f33ef4963c87a72b33f1c8d0fa`
`ddim_rope_50`	`40000`	`d4c50efe6711cf387e3801623bee554366908613346f6c49c978e99a65a9f1aa`	`4793c5d2f784560eeba511b534466aacd5d91738b3d1ae3403e7597dc3dd7933`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`043fd62974cb6cd1de5f3d3671d831122922b62cfa6523166bdf2eefb1112b41`
`ddim_rope_50`	`50000`	`fcb539e3a2c269cd964756b820697be031ca689fbee8d30906fdffba6ab78dd1`	`4793c5d2f784560eeba511b534466aacd5d91738b3d1ae3403e7597dc3dd7933`	`97c1bcc16b591683c3a554b8add91778631ad78918ccea25fa73c8615e2f1b18`	`326cd883f40e5e27d0c9dcc1fbdcfb5cdf4b9a3ed14732dfedc043ae725483b6`

Published Layout

baseline/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddpm_20/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddpm_50/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddpm_rope/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddim_20/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddim_50/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddim_rope_20/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt
ddim_rope_50/
  checkpoints/
    40000/
      model.safetensors
      config.json
      ramen_stats.pt
    50000/
      model.safetensors
      config.json
      ramen_stats.pt

Provenance

The source experiment metadata lives in run_logs/coffee_capsules_category1_core_diffusion/index.md and the per-variant logs in run_logs/coffee_capsules_category1_core_diffusion/.

These published artifacts were uploaded directly from the original training outputs on scratch rather than from a separate local staging copy.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support