recipe_wd_t1_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
Weight decay sized to a one-epoch memory timescale helps when data repeats.
What we learned
Yes: ex-Gita 0.6971, 2.4% better than the reference.
Scores
Bits per byte, lower is better; change against the parent run, RECIPE R0.
ex-Gītā (headline)
0.6971
bits per byte
−2.5%ex-Gītā, clean_v1
0.7225
bits per byte
−2.7%Pooled, all five sets
0.6877
bits per byte
−2.5%Validation split
0.6767
bits per byte
−1.8%| Test set | RECIPE WD τ1 | RECIPE R0 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.684 | 0.6993 | −2.2% |
| Bhagavad-gītā (memorisation) | 0.4403 | 0.4431 | −0.6% |
| Out of domain | 0.7007 | 0.7186 | −2.5% |
| Prose | 0.6619 | 0.6755 | −2.0% |
| Vedic (Ṛgveda) | 1.0863 | 1.1793 | −7.9% |
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
- 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