recipe_arch_d125m
d125m (103.4M parameters) on plus_all v2 subset (4 passes) (28.3M words, 4.02 passes), trained on RTX 3060 12GB (Home GPU). Started 5 Oct 2026, ended 5 Oct 2026.
Hypothesis
Value embeddings, a logit soft-cap and document-boundary attention masking help.
What we learned
Lose: 3.9% worse than its layout control. The mask cuts each short evaluation item off from its neighbours; with it off at evaluation the model ties.
Scores
Bits per byte, lower is better; change against the parent run, RECIPE R0docs.
ex-Gītā (headline)
0.7542
bits per byte
+3.9%ex-Gītā, clean_v1
0.804
bits per byte
+5.7%Pooled, all five sets
0.7473
bits per byte
+4.3%Validation split
0.6568
bits per byte
+0.7%| Test set | RECIPE ARCH | RECIPE R0docs (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.8241 | 0.7478 | +10.2% |
| Bhagavad-gītā (memorisation) | 0.5664 | 0.4682 | +21.0% |
| Out of domain | 0.7366 | 0.7212 | +2.1% |
| Prose | 0.7251 | 0.6873 | +5.5% |
| Vedic (Ṛgveda) | 1.4643 | 1.3191 | +11.0% |
Curves
Training loss
Cross-entropy per token, by step. The first few percent of the run, far higher, run off the top; hover or the table has every value.
Held-out bits per byte
The validation split, evaluated during training, by step. Lower is better.
Throughput
Tokens per second, by step.
Model
- Preset
- d125m
- Parameters
- 103,385,868
- Outside embeddings
- 84,953,868
- Layers · heads · width
- 12 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- plus_all v2 subset (4 passes)
- Words
- 28,300,000
- Training tokens
- 90,886,950
- Passes
- 4.02 passes
- Tokens seen
- 364,953,600
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,425 / 7,425
- GPU hours
- 3.46
- Cost
- —
- Spot restarts
- —
Lineage
homed125m