Upload PHASE6_HETERO_POS.md with huggingface_hub

PHASE6_HETERO_POS.md

@@ -0,0 +1,87 @@
+# Phase 6 — heterogeneous positional bias: content–position balance ties, the temperature knob hurts
+
+A final orthogonal probe (2026-06-23) closes the applied attention arc by testing
+the *only* axis the earlier bake-offs had not: per-head **content↔position
+balance**. The verdict re-confirms the Phase-5 selection — **learnable-ALiBi (one
+learnable linear slope per head, nothing else) is the hypothesis to ship.**
+
+## Why this experiment
+
+The GPU bake-off winner, learnable-ALiBi, learned per-head slopes spanning ~125×
+(0.68 → 0.005), with the flattest heads going essentially NoPE-like. So its +8%
+was never "a better distance curve" — it was the model **self-organizing into
+sharp-local heads plus near-global content-only heads**. Every variant that added
+more *distance shape* (Calabi–Yau curve, log-curvature, Fourier features) overfit
+and lost at scale. The untried, orthogonal axis is to make that local↔global
+mixture **explicit**: give each head a learnable content scale `c_h` so it can
+choose how much weight the content score `q·k` carries relative to the positional
+penalty `−a_h·d`:
+
+```
+score_h(i,j) = c_h · (q_i · k_j)/√D  −  a_h · (i − j)        [content_balance]
+```
+
+**Hypothesis:** making the local↔NoPE mixture explicit (learnable slope `a_h`
+*and* learnable content scale `c_h`) should modestly improve perplexity and — the
+real prize — improve **length extrapolation**, without the overfit that killed
+"more shape". A per-head softmax temperature `τ_h` was added as an extra-knob
+ablation. **Pre-committed KILL:** if `content_balance` gives no extrapolation gain
+over `alibi_learn`.
+
+Schemes: `alibi_fixed` (baseline), `alibi_learn` (the bake-off winner / control),
+`nope` (content-only sanity), `content_balance` (the hypothesis), `cb_softmax_temp`
+(content_balance + per-head temperature).
+
+## GPU result (NVIDIA L4, 2026-06-23, full preset)
+
+Trained from scratch on real WikiText-2 (`Salesforce/wikitext`, wikitext-2-raw-v1),
+d_model 512 / 8 layers / 8 heads, ctx 512, 4000 steps, batch 24, γ-L2 1e-3, val
+early-stop. **Primary metric: validation perplexity at 1×/2×/4×/8× the training
+context (length extrapolation; lower is better).**
+
+| mode | 1× (512) | 2× (1024) | 4× (2048) | 8× (4096) |
+|---|---|---|---|---|
+| `alibi_fixed` | 3.705 | 3.623 | 3.577 | 3.513 |
+| **`alibi_learn`** (control / winner) | 3.651 | 3.559 | 3.500 | 3.429 |
+| `nope` | 10.290 | 10.845 | 11.571 | 12.311 |
+| **`content_balance`** (hypothesis) | 3.642 | 3.548 | **3.488** | **3.428** |
+| `cb_softmax_temp` (+ temperature) | 3.732 | 3.659 | 3.599 | 3.548 |
+
+`content_balance` vs `alibi_learn` @4×: **+0.35%** (and +0.17% @8×).
+
+## Verdict — KILL the extra knobs; learnable-ALiBi stands
+
+1. **`content_balance` ties `alibi_learn` within noise** (+0.35% @4×, +0.17% @8×).
+   The learnable content scales settled near 1.0 with a mild upward drift toward
+   the flatter/global heads (`c` ≈ 0.79 → 0.93 across the slope-ordered heads) —
+   the model *does* want global heads to lean a little more on content, but the
+   effect is marginal and does not clear a meaningful margin. **KILL per the
+   pre-committed criterion** (no extrapolation gain over `alibi_learn`).
+2. **The softmax-temperature knob *regresses*** — `cb_softmax_temp` is worse than
+   plain `alibi_learn` at every length (3.732 vs 3.651 @1×). This is the recurring
+   Phase-4/5 pattern once more: **extra parameters hurt at scale.** Notably the
+   *local* screen (d128/L3, 800 steps) had crowned `cb_softmax_temp` the best
+   scheme — another instance of a short, small screen rewarding the highest-capacity
+   variant that then loses when properly trained.
+3. **`nope` collapses and gets *worse* with length** (10.3 → 12.3), the expected
+   sanity check that the positional penalty is doing real work.
+
+## What this adds to the arc
+
+Phase 6 was the last orthogonal axis left to try — not *more distance shape*
+(Phases 4–5 closed that), but *content/position mixing*. It lands in the same
+place: **adaptability beats added structure, but only up to the single learnable
+slope.** Once each head can pick its own slope, an explicit content-balance knob is
+redundant (the model already realizes the local↔global split through the slope
+alone) and a temperature knob is actively harmful.
+
+> **Selected hypothesis (final):** a **learnable per-head linear positional bias**
+> (learnable-ALiBi) — one slope per head, no log-curvature, no Calabi–Yau sequence,
+> no content scale, no temperature. It is the robust GPU-scale winner across every
+> bake-off (Phases 4, 5, 6), passes the §5 gate at +8.1% over the best baseline,
+> and extrapolates flat where learned-absolute positions collapse.
+
+Artifacts: `4_ai_hardware_attention/cy_sieve_hetero_pos.py` (experiment),
+`gpu_phase_runs/hetero_pos_gpu_20260623.json` + `.log` (GPU run),
+`gpu_phase_runs/hetero_pos_local_20260623.json` (local screen). See also
+`PHASE5_FINAL_BAKEOFF.md` and `CYSIEVE_JOURNEY.md`.