f7_slp1_uni8k_d350m_plus_clean_2x
d350m (318.4M parameters) on plus_clean slice (448.9M words, 1.84 passes), trained on A100 40GB (Google Cloud, spot). Started 2 Oct 2026, ended 3 Oct 2026.
Hypothesis
A 350M model on the same data and steps beats the 125M model.
What we learned
Yes: ex-Gita 0.5720, 5.3% better, Vedic 10% better. Our first cloud run (spot A100). Released as sansar-350m v0.1.0.
Scores
Bits per byte, lower is better; change against the parent run, F6-clean-2x.
ex-Gītā (headline)
0.572
bits per byte
−5.3%ex-Gītā, clean_v1
0.5937
bits per byte
−5.1%Pooled, all five sets
0.5568
bits per byte
−5.9%Validation split
0.5102
bits per byte
−10.9%| Test set | F7 | F6-clean-2x (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.5757 | 0.6052 | −4.9% |
| Bhagavad-gītā (memorisation) | 0.1552 | 0.279 | −44.4% |
| Out of domain | 0.573 | 0.6052 | −5.3% |
| Prose | 0.5577 | 0.5859 | −4.8% |
| Vedic (Ṛgveda) | 0.7052 | 0.7833 | −10.0% |
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
- d350m
- Parameters
- 318,424,064
- Outside embeddings
- 302,040,064
- Layers · heads · width
- 24 · 16 · 1,024
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- plus_clean slice
- Words
- 448,900,000
- Training tokens
- 1,444,720,245
- Passes
- 1.84 passes
- Tokens seen
- 2,664,628,224
Compute
- GPU
- A100 40GB
- Where
- Google Cloud (spot)
- Steps
- 54,212 / 54,212
- GPU hours
- 14.01
- Cost
- $18.22
- Spot restarts
- 2
Lineage
gcpa100spotd350m