recipe_r0_d125m
d125m (97.2M parameters) on plus_all v2 subset (4 passes) (28.3M words, 4 passes), trained on RTX 3060 12GB (Home GPU). Started 4 Oct 2026, ended 4 Oct 2026.
Hypothesis
Reference: the F4 recipe in the 700M run's data regime (four passes over a fixed subset).
What we learned
ex-Gita 0.7152. Absolute numbers are not comparable with the F series (small subset).
Scores
Bits per byte, lower is better; change against the parent run, F4.
ex-Gītā (headline)
0.7152
bits per byte
+15.6%ex-Gītā, clean_v1
0.7427
bits per byte
Pooled, all five sets
0.7052
bits per byte
+15.9%Validation split
0.6892
bits per byte
+16.9%| Test set | RECIPE R0 | F4 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6993 | 0.6204 | +12.7% |
| Bhagavad-gītā (memorisation) | 0.4431 | 0.3271 | +35.5% |
| Out of domain | 0.7186 | 0.6158 | +16.7% |
| Prose | 0.6755 | 0.5974 | +13.1% |
| Vedic (Ṛgveda) | 1.1793 | 1.0098 | +16.8% |
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
- 91,236,966
- Passes
- 4 passes
- Tokens seen
- 364,953,600
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,425 / 7,425
- GPU hours
- 3.39
- Cost
- —
- Spot restarts
- —
Lineage
homed125m