ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints
9f1076a verified |
Download docs/FIELD_IGNORANCE.md from thefinalboss/atom-ai: direct link, hf CLI and curl.
- Browser
- Download file 2.18 kB
-
https://huggingface.co/thefinalboss/atom-ai/resolve/main/docs/FIELD_IGNORANCE.md
- Command line
-
hf download hf://thefinalboss/atom-ai/docs/FIELD_IGNORANCE.md
-
curl -L -o FIELD_IGNORANCE.md https://huggingface.co/thefinalboss/atom-ai/resolve/main/docs/FIELD_IGNORANCE.md
2.18 kB
| # Field-ignorance loss | |
| _Numbers only. English._ | |
| ## Formula | |
| ``` | |
| L_ign = ReLU( cos( ℓ(α), ℓ(sg[α_bar]) ) − m ) | |
| ``` | |
| - `ℓ(·)` = surface byte logits (same head as CE), from field `α` vs from stopgrad `α_bar` | |
| - `α_bar` = sliding mean of the rolling α-bank of **other** prompts (not zero, not same-prompt) | |
| — or, with `--field-ignorance-prompt-bank`, mean of a **fixed distinct-prompt** α bank | |
| - `m` = **0.85** | |
| - weight = **0.08** | |
| Purpose: penalize a surface that maps any non-zero α to one shared logit template. | |
| ## Flags | |
| | Flag | Default | Notes | | |
| |------|---------|-------| | |
| | `--field-ignorance-weight` | `0.08` | `0` disables | | |
| | `--field-ignorance-margin` | `0.85` | hinge margin | | |
| | `--field-ignorance-ablate-shared` | off | ℓ(α_bar) gets `atom_r`/`persistence` = None (zeros); ℓ(α) live | | |
| | `--field-ignorance-prompt-bank` | off | α_bar from probe-style fixed prompts, not train ticks | | |
| Wired in `AtomNativeModel._field_auxiliary_losses` next to field contrast. | |
| ## Merge diagnostics (read-only) | |
| When the merge path runs, log: | |
| - `max_phase_coherence` — max pairwise phase-coherence **before** threshold filter | |
| - `n_pairs_above_energy_floor` — pairs that clear the energy-floor checks | |
| Thresholds `phase_coherence_threshold` / `merge_energy_floor` are **unchanged**. | |
| ## Comparison (smokes +2500 from same src @ 2102500) | |
| | run | logits cos (before→after) | α cos (before→after) | ign (final) | mph | merges | | |
| |-----|---------------------------|----------------------|-------------|-----|--------| | |
| | baseline L_ign | 0.992294→**0.991285** | 0.720524→0.698694 | **0.150** | 0.500 | 0 | | |
| | ablate-shared | 0.992294→**0.991780** | 0.720524→0.698509 | **0.150** | 0.500 | 0 | | |
| | prompt-bank | 0.992294→**0.991156** | 0.720524→0.698703 | **0.000** | 0.500 | 0 | | |
| | both | 0.992294→**0.992358** | 0.720524→0.698481 | **0.000** | 0.500 | 0 | | |
| **Success bar:** logits drop vs ~0.991 without killing α (~0.70). **Not met** on any of the three mechanism smokes. | |
| Detail: `docs/LIGN_SMOKE.md`, `docs/LIGN_ABLATION.md`, `docs/LIGN_PROMPT_BANK.md`, `docs/LIGN_ABLATION_PLUS_BANK.md`, `docs/artifacts/lign_readout_series/SUMMARY.md`. | |