---
library_name: transformers
pipeline_tag: text-generation
tags:
- llama
- causal-lm
- bfloat16
---

# epe-3p-smollm-1p7b-100B-20n-2048sl-960gbsz

Converted Hugging Face base checkpoint from the Model Raising EPE (Ethics-by-Pretraining) pretraining run.

## Details

- Architecture: `LlamaForCausalLM`
- Base model size: `1.7B`
- Precision on disk: `bfloat16`
- Tokenizer: extended SmolLM2 tokenizer with 36 additional special tokens (`<assistant>` + 35 `<charter_X.Y>` tokens), vocab size 49280

## EPE Pretraining

This model was pretrained with on-the-fly reflection insertion using the `reflection_3p` column from the annotated sidecar dataset. The training augments standard autoregressive NTP with:

1. **Reflection insertion**: reflections are inserted into annotated documents at inference time; the model predicts the reflection tokens using CE loss
2. **Constitution predictor**: a multi-label BCE loss at the `<assistant>` token position trains the model to predict 35 charter items
3. **Attention masking**: post-reflection context tokens are blocked from attending to the reflection region
4. **Position aliasing**: post-reflection context tokens alias back to the same RoPE positions as pre-reflection context, making inference (without reflections) positionally equivalent to training

## Chat Templates

Two named chat templates are provided:

| Name | Use case |
|------|----------|
| `default` | Standard SFT — plain `assistant` role token |
| `epe` | Activates constitution head — uses `<assistant>` (token 49152) at start of assistant turns |

```python
tok.apply_chat_template(messages, chat_template="default")  # standard
tok.apply_chat_template(messages, chat_template="epe")      # constitution head active
```

## SFT Notes

- Always use the bundled tokenizer (vocab size 49280); the original SmolLM2 tokenizer (49152 tokens) will mismatch embeddings for IDs 49152–49279
- `vocab_size=49280` is set in `config.json`