Udgam English 125M v0.1.0
A small English language model trained from random weights, with a vocabulary fitted on its training text. We share the base and conversation variants, the recipe, and the results — including the missed performance target.
The result
A completed experiment. A missed performance target.
The performance target was missed. Both variants scored below Sarvam-1 and both SmolLM2 comparisons, and failed every predefined task floor and the overall floor. We publish the negative results so others can inspect the method and build on the work.
What we trained
- Parameters
- 125,829,314
- Initialization
- Random weights; no pretrained checkpoint
- Vocabulary
- 16,384 tokens, fitted on training text only
- Context
- 2,048 tokens
- Pretraining
- 2.03 billion prediction targets
- Conversation tuning
- One pass over 7,237 prepared OASST2 examples
- Format
- Native PyTorch with an offline inference entrypoint
The architecture and training implementation build on the open-source nanochat project. Public training sources include FineWeb-Edu, FinePhrase and Nemotron fact-seeking text. Some text was generated by other models; their identities and source terms are recorded in the recipe.
Measured results
| Model | Accuracy |
|---|---|
| Udgam English 125M · base | 45.72% |
| Udgam English 125M · conversation | 46.40% |
| Sarvam-1 | 68.08% |
| SmolLM2-135M | 51.76% |
| SmolLM2-360M | 57.40% |
Macro accuracy on a fixed 2,500-question suite: 500 items each from ARC-Challenge, HellaSwag, PIQA, BoolQ and WinoGrande. These scores describe this protocol and these checkpoints, not general assistant quality.
Both variants also passed 0 of 24 strict automatic instruction checks. These are limited diagnostics; the separate 12 human-rubric items per variant remain unscored. The full report retains uncertainty estimates, failed criteria and per-item outcomes.
Read the full evaluationRun, inspect and adapt
The base variant completes text. The conversation variant uses native chat role markers. Both run locally with the included tokenizer. The release includes a guide for fine-tuning on your own prompt-and-response examples.
Verify the published checksums before use. This native package is not a Transformers, GGUF or quantized release. New fine-tunes do not inherit these benchmark scores.
Read the adaptation guideScope and limitations
Intended for small-model research and experimentation. Output can be incorrect, repetitive, biased or fail instructions. The measured results do not support a reliable general assistant or use in high-stakes decisions. Multilingual quality, coding ability and long-context behavior have not been established.
Udgam-owned weights, tokenizer, code and documentation use MIT terms with upstream notices retained. Third-party datasets keep their own licenses. The release shares the acquisition recipe, not a raw third-party corpus or comparison-model weights. Full historical training requires the pinned inputs and environment; it is not a one-command service.