CetinLM Base-v1
From-scratch language-model pretraining · public model card · active research
→ View the latest CetinLM research & training metrics
Model card
CetinLM Base-v1 is a ~1.18B-parameter decoder-only language model trained from scratch with the custom CetinTokenizer-v1 tokenizer.
This page describes the public Base-v1 research lineage. The model is still in the base-pretraining stage; instruction following, assistant behavior and later post-training capabilities are separate stages and should not be inferred from pretraining loss alone.
| Model ~1.18B decoder-only parameters. |
Tokenizer CetinTokenizer-v1 · 48K vocabulary. |
Context 2K active Base-v1 training context. |
The frozen Base-v1 corpus contains approximately 11.39B unique training tokens. Processed-token count measures training exposure and is not the same thing as unique corpus size.
Latest metrics live on the Research page
CetinLM is actively training, which means the latest token count, validation loss, perplexity, per-dataset measurements and sampled generation-health results continue to change.
To avoid leaving stale numbers scattered across model cards and repositories, the canonical current metrics are maintained here:
CetinLM Research → cetinlm.meforcetechnology.com/research
Historical measurements below remain intentionally preserved as a public reference snapshot.
Archived public snapshot — 5.15B
| Signal | Archived value |
|---|---|
| Processed tokens | 5.15B |
| Validation loss | 2.496678 |
| Perplexity | 12.142 |
| Checkpoint status | NEW BEST at snapshot |
EOS P(EOS@end) |
0.3994 |
| EOS top-1 / top-5 | 45.0% / 75.0% |
| Median EOS rank | 2.0 |
This is an archived reference point, not the current training state. For the current checkpoint, use the live Research page above.
Reference training trajectory
| Milestone | Val loss | PPL |
|---|---|---|
| 4.80B | 2.513649 | 12.350 |
| 4.85B | 2.504830 | 12.241 |
| 4.90B | 2.502269 | 12.210 |
| 4.95B | 2.501286 | 12.198 |
| 5.00B | 2.500880 | 12.193 |
| 5.05B | 2.503461 | 12.225 |
| 5.10B | 2.502036 | 12.207 |
| 5.15B | 2.496678 | 12.142 |
Across the archived 4.80B → 5.15B window, held-out validation loss moved from 2.513649 → 2.496678.
Sampled generation-health reference
At 5.00B, a separate 1,000-generation web-balanced-v1 evaluation recorded:
- loop incidents: 5/1000 (0.500%)
- severe loops: 0/1000 (0.000%)
- repetition burden: 0.020% — 20 / 101,597 generated tokens
- EOS stops: 393/1000
- length limit: 602/1000
- average generated tokens: 101.6
The 25-case RAW GREEDY STRESS PROBE is intentionally harsher and is not presented as a user-facing loop rate.
For newer sampled-behavior runs, see CetinLM Research.
Training-stage boundary
Base-v1 is the pretrained foundation, not the finished assistant.
The current model card should therefore be read as evidence about:
- language-model pretraining progress;
- held-out validation behavior;
- dataset-level validation trends;
- EOS boundary health;
- mechanical generation diagnostics.
It should not be read as a final claim about instruction following, reasoning, coding, factuality, safety, tool use or assistant quality. Those belong to later post-training and dedicated evaluations.
CetinLM Live
CetinLM Live is a separate interactive runtime layer developed around the model stack. Public work includes realtime speech, synchronized text/audio interaction, interruption handling, local voice processing and persistent interaction research.
Live/runtime development does not alter Base-v1 checkpoint accounting or the frozen training lineage.
Speech-runtime licensing
Optional Live speech uses separately licensed third-party software. Resemble AI Chatterbox Multilingual V3 is used for text-to-speech under its upstream MIT License. CetinLM does not claim ownership of Chatterbox.
Public third-party provenance and license notices are maintained with the CetinLM GitHub public research archive.
Research links
- Latest metrics & research telemetry: cetinlm.meforcetechnology.com/research
- Project website: cetinlm.meforcetechnology.com
CetinLM · build from scratch, measure continuously, keep the evidence public.