recipe_r0docs_d125m
d125m (97.2M 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
Layout control: lines of one text joined into documents.
What we learned
ex-Gita 0.7262, 1.7% worse than the per-line layout. The reference for ARCH.
Scores
Bits per byte, lower is better; change against the parent run, RECIPE R0.
ex-Gītā (headline)
0.7262
bits per byte
+1.5%ex-Gītā, clean_v1
0.7607
bits per byte
+2.4%Pooled, all five sets
0.7167
bits per byte
+1.6%Validation split
0.6524
bits per byte
−5.3%| Test set | RECIPE R0docs | RECIPE R0 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7478 | 0.6993 | +6.9% |
| Bhagavad-gītā (memorisation) | 0.4682 | 0.4431 | +5.7% |
| Out of domain | 0.7212 | 0.7186 | +0.4% |
| Prose | 0.6873 | 0.6755 | +1.7% |
| Vedic (Ṛgveda) | 1.3191 | 1.1793 | +11.9% |
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
- 97,241,856
- Outside embeddings
- 84,953,856
- 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
- 3.38
- Cost
- —
- Spot restarts
- —
Lineage
homed125m