---
language: en
tags:
  - loracle
  - lora-oracle
  - mechinterp
  - llama-3-3-70b
license: mit
base_model: meta-llama/Llama-3.3-70B-Instruct
---

# llamacle_v6_clean - Loracle on Llama-3.3-70B (1-epoch pretrain, step 1875)

Loracle = a model that reads LoRA weight deltas and describes the behavioral
change without ever running the fine-tuned model. This is the Llama-70B
version, end of 1-epoch pretrain on 22.5k diverse per-org LoRAs (oneq dataset).

## Stack
- **Base**: meta-llama/Llama-3.3-70B-Instruct (frozen, bf16)
- **Direction tokens**: SVD top-16 x 80 layers x 7 mag-7 sides = `[8960, 8192]` bf16
- **Interpreter LoRA**: rank=256, alpha=32, rsLoRA, on the frozen base
- **Encoder**: norm-match injection at layer 1
- **Trainer**: FSDP2, AdamW fp32 master params (bf16 weights), constant LR (no warmup)

## Training config
- 22,500 train orgs / 2,500 holdout
- 1875 opt steps, effective batch 12 (6 ranks x bs=1 x ga=2)
- max_length=9500, n_direction_tokens=8960
- ~6.5 hours wall on 6xB200

## Final metrics (step 1875)
- val_loss = 1.7042
- cross-LoRA gap = +0.4434 (matched=1.4506, crossed=1.8940)

## Eval-loss progression across log-spaced ckpts

| step | val_loss | cross_lora_gap |
|---|---|---|
| 47 | 3.30 | +0.05 |
| 79 | 2.24 | +0.16 |
| 134 | 2.07 | +0.24 |
| 228 | 1.94 | +0.35 |
| 386 | 1.85 | +0.32 |
| 654 | 1.78 | +0.42 |
| 1107 | 1.74 | +0.43 |
| 1875 | 1.70 | +0.44 |

## Notes
- bf16 base + fp32 LoRA params (avoids bf16 underflow on Adam first-step)
- This is a pretrain checkpoint; SFT+RL post-training to follow.