---
license: cc-by-nc-4.0
tags:
- dna
- enhancer-generation
- mdlm
- autoregressive
- ntv3
- unconditional-baseline
---

# NTV3-MDLM-8M unconditional enhancer generator (T1)

Unconditional NTv3 generation baseline for the brain 7-cell
enhancer task. There is **no cell-type signal** at training or
inference time, so this model learns ``p(enhancer)`` and emits a
single mode-collapsed enhancer for every gold input regardless of
the requested cell type. Useful as a lower bound for cell-type
specificity (see the MDLM/AR rows in
[explcre/biomodel_reasoning_calling_study2](https://github.com/explcre/biomodel_reasoning_calling_study2)
``regureasoner_loop/docs/EXPERIMENTS.md``).

Trained on `train.enhancer_generation.strat7c.n35k` (35k rows
stratified across 7 brain cell types). MDLM-8m / MDLM-650m use the
weighted-CE diffusion loss (1/t schedule); AR-650m uses standard
next-token CE.

## Files

```
best.pt                    # EMA / best-checkpoint state dict
metrics.json               # 28-column eval against gold n7k held-out test
training_log.jsonl         # per-step train metrics
manifest.json              # run config snapshot
predictions_preview.jsonl  # first 200 sampled rows (for spot-check)
```

## Headline eval (held-out gold n7k)

| parse_rate | argmax_acc | on_target | off_target | specificity | FID |
|---:|---:|---:|---:|---:|---:|
| 1.000 | 0.163 | -18.655 | -18.905 | 0.251 | 18.420 |


Random argmax baseline over 7 cells = 0.143; the unconditional
baselines all sit at chance because the model has no signal to
differentiate cell types. FID is misleadingly low (the average
enhancer is statistically plausible) but cell-type recall is
absent — see the cell-type-conditioned MDLM repos for the
conditional fix.

## Citation
See the parent repo's `EXPERIMENTS.md` for the full per-baseline
comparison table.