oracle_baseline_noVE_d24_fp8

What: nanochat d24 base-model checkpoints (every 2000 steps + final step 8352; gzipped model_*.pt.gz + meta_*.json, optimizer state included) — negative control, no oracle injection, value embeddings ZEROED + FROZEN (NO_VALUE_EMBEDS=1).

Trained on: karpathy/climbmix-400b-shuffle (first ~274 shards), 8.758B tokens = compute-optimal ratio 12, seed 1337, bf16 + fp8 GEMM (tensorwise), 8xH100. Script: nanochat/runs/oracle_runs/no_value_embeds.sh in oracle-encodings.

Result: CORE 0.2711 at step 8352.

Used for: the exact-match control for the Stage-7 injected run (which also trains no-VE, same seed/steps/data); value embeddings are removed because they are a confound for the injection. Sibling VE baseline: oracle_baseline_d24_fp8. The first injection store (built from a 162-output joint encoder) was deleted 2026-07-09 as a design error; treatment runs will use per-layer stores (one gemma layer per nanochat model) when rebuilt.

Status: active (frozen reference).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support