Files

71 lines
2.6 KiB
Markdown

# Attention Residuals train-time forward intervention
This directory implements preregistered protocol
`llm-atlas-k3-attnres-forward-training-v1`:
- `research/K3_ATTNRES_FORWARD_TRAINING_SCOPING.md`
- `research/K3_ATTNRES_FORWARD_TRAINING_PROTOCOL.md`
- `research/K3_ATTNRES_FORWARD_TRAINING_GROK_REVIEW.md`
- `research/K3_ATTNRES_FORWARD_TRAINING_IMPLEMENTATION_REVIEW.md`
- `research/K3_ATTNRES_FORWARD_TRAINING_AUDIT.md`
It is a depth-32 reduced Block AttnRes architecture ablation. It is not a
Kimi-K3 checkpoint forward pass and does not claim to recover unpublished
Figure 5 telemetry.
## Frozen environment
```text
Python /home/wuyang/.pyenv/versions/3.10.14/envs/navi-router-cu128/bin/python
PyTorch 2.11.0+cu128
GPU NVIDIA GeForce RTX 5090
CUBLAS_WORKSPACE_CONFIG=:4096:8
maximum concurrency 2
```
## Pre-result gates
The checked-in gate artifacts must pass before formal output:
```bash
CUBLAS_WORKSPACE_CONFIG=:4096:8 \
/home/wuyang/.pyenv/versions/3.10.14/envs/navi-router-cu128/bin/python \
experiments/k3/attnres_forward/verify.py step-zero \
--cache-dir /home/wuyang/.cache/llm-atlas/k3-attnres-gradient-scale-v1 \
--parent-manifest experiments/k3/attnres_gradient/manifest.json \
--study-manifest experiments/k3/attnres_forward/manifest.json \
--output experiments/k3/attnres_forward/results/gates/step-zero.json
```
`learned_reference` is smoke-only. Its 20-step result is compared with a
fresh parent Round 05 smoke using `verify.py smoke-compare`.
## Formal matrix
```bash
/home/wuyang/.pyenv/versions/3.10.14/envs/navi-router-cu128/bin/python \
experiments/k3/attnres_forward/run_matrix.py \
--python /home/wuyang/.pyenv/versions/3.10.14/envs/navi-router-cu128/bin/python \
--cache-dir /home/wuyang/.cache/llm-atlas/k3-attnres-gradient-scale-v1 \
--parent-manifest experiments/k3/attnres_gradient/manifest.json \
--study-manifest experiments/k3/attnres_forward/manifest.json \
--output-dir experiments/k3/attnres_forward/results/raw \
--phase all \
--concurrency 2
```
This runs 12 formal cells and one full replay. The analyzer reads all cells,
the frozen historical paired references, and generates the only authoritative
status, interaction map, and website compact artifact.
The checked-in Round 08 release contains:
- 13 raw results under `results/raw/`;
- `reproduction.json` with the exact primary scientific-payload hash;
- aggregate / compact website data under `src/data/`;
- a frozen-data checker and real-Chrome five-view regression in `scripts/`.
The established status is deliberately scoped to this reduced protocol. It is
not a real Kimi-K3 checkpoint result or a reproduction of unpublished Figure
5(c) telemetry.