Skip to content

Releases: Bl4ckd09/nanochat-on-rtx4070

Archived Base Export d24_asp48_track s615173

Choose a tag to compare

Archived base export for d24_asp48_track at step 615173.

Contents:

  • full export directory with model_615173.pt, optim_615173_rank0.pt, meta_615173.json, and SHA256SUMS.txt

Download all parts and reconstruct with:

cat d24_asp48_track_s615173_2026-03-13_154421.tar.gz.part-* > d24_asp48_track_s615173_2026-03-13_154421.tar.gz
sha256sum -c SHA256SUMS_split.txt

Best observed unstable manual checkpoint: manual_reasoning_v1_seed42_best_observed_2026-04-02_0435

Choose a tag to compare

Exported Best-Observed Unstable Manual Checkpoint

This release packages the strongest single-run result from the manual-curation branch.

  • Run: reasoning_manual_v1_seed42_2026-04-01_1315_s768_gc
  • Base: d24_asp48_track @ 820230
  • Checkpoint step: 300
  • Status: best_observed_but_not_promoted

Confirmed metrics

  • GSM8K pass@8: 6.70%
  • MMLU: 27.60%
  • SpellingBee: 0.00%

Why it is not promoted

  • This was the strongest single-run reasoning result observed on the project.
  • The paired seed in the same sweep failed the promotion rule, so the recipe remained unstable.
  • Official project status still keeps mixv2_s768_300 as the provisional champion.

Release asset layout

This archive was split into multiple release assets because GitHub release assets must be under 2 GiB each.

Assets to upload:

  • manual_reasoning_v1_seed42_best_observed_2026-04-02_0435.tar.gz.part-00
  • manual_reasoning_v1_seed42_best_observed_2026-04-02_0435.tar.gz.part-01
  • SHA256SUMS_split.txt

Reconstruction after download:

cat manual_reasoning_v1_seed42_best_observed_2026-04-02_0435.tar.gz.part-* > manual_reasoning_v1_seed42_best_observed_2026-04-02_0435.tar.gz
sha256sum manual_reasoning_v1_seed42_best_observed_2026-04-02_0435.tar.gz

Champion checkpoint: d24_r32_adamw_partial_fr20_mixv2_s768_s300_champion_2026-03-30_190948

Choose a tag to compare

Exported Champion Checkpoint

This release packages the current provisional chat-side champion from the consumer-GPU branch.

  • Run: d24_r32_adamw_partial_fr20_mixv2_2026-03-29_1040_s768_gc
  • Base: d24_asp48_track @ 820230
  • Checkpoint step: 300
  • Status: current_champion_provisional

Confirmed metrics

  • GSM8K pass@8: 4.60%
  • MMLU: 27.40%
  • SpellingBee: 0.39%
  • Best SFT val bpb: 0.6981266376087805

Release asset layout

This archive was split into multiple release assets because GitHub release assets must be under 2 GiB each.

Assets to upload:

  • d24_r32_adamw_partial_fr20_mixv2_s768_s300_champion_2026-03-30_190948.tar.gz.part-00
  • d24_r32_adamw_partial_fr20_mixv2_s768_s300_champion_2026-03-30_190948.tar.gz.part-01
  • SHA256SUMS_split.txt

Reconstruction after download:

cat d24_r32_adamw_partial_fr20_mixv2_s768_s300_champion_2026-03-30_190948.tar.gz.part-* \
  > d24_r32_adamw_partial_fr20_mixv2_s768_s300_champion_2026-03-30_190948.tar.gz
sha256sum d24_r32_adamw_partial_fr20_mixv2_s768_s300_champion_2026-03-30_190948.tar.gz

Notes

  • This is the strongest single confirmed run on the current d24/r32 chat branch.
  • It remains provisional because later exact replication and seed sweeps did not reproduce it reliably.
  • The original unsplit archive remains available locally in ~/nanochat-learn/exports/.
  • GitHub docs currently state each release file must be under 2 GiB.

Base Champion d24_asp48_track r32 s820230

Choose a tag to compare

Base Champion Export

Stable base champion for the consumer-GPU nanochat branch.

Model

  • model tag: d24_asp48_track
  • checkpoint step: 820230
  • status: stable base champion

Base metrics

  • train bpb: 0.801367
  • val bpb: 0.919863
  • CORE: 0.1514

Contents

  • base_champion_r32_820230_full_2026-03-31.tar.gz.part-*
    • full base export
    • includes model, optimizer, metadata, checksums, and README
  • base_champion_r32_820230_weights_only_2026-03-31.tar.gz.part-*
    • weights-only base export
    • includes model, checksums, and README
  • SHA256SUMS_split.txt

Notes

  • Files are split because GitHub Release assets have a per-file size limit.
  • Reconstruct locally before extracting.
  • This release is separate from the chat champion release because base and chat exports serve different roles.