Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,109 @@
+# hipfire-calib-data
+
+Native BF16-teacher calibration artifacts (`.calib.hfq`) and fixed KLD
+reference tensors (`.ref_*.bin`) produced by the hipfire quantization-quality
+campaign. One folder per architecture.
+
+Produced on an AMD MI300X (`gfx942`, CDNA3). **All scoring was performed
+elsewhere** (`gfx1201`); the MI300X was a producer only, because uncalibrated
+MQ4 is degenerate on gfx942.
+
+## `.calib.hfq` — HFQM calibration package
+
+```
+header 32B: magic "HFQM" | version u32=1 | arch_id u32 | n_entries u32
+            | metadata_offset u64 | data_offset u64
+metadata:   self-delimited JSON blob
+index:      n_entries u32 | per-entry { name_len u16 | name | quant_type u8
+            | n_dims u8 | shape[n_dims u32] | group_size u32 | data_size u64 }
+payloads:   concatenated, in index order, from data_offset
+```
+
+Each tensor contributes two entries: `<name>.hessian` and `<name>.imatrix`.
+A 496-tensor model therefore has `n_entries = 992`.
+
+### Hessian storage: `Bf16TrilDiagF32` (quant_type 130)
+
+Per Hessian, `K*4 + K*(K-1)/2 * 2` bytes:
+
+- **diagonal** — exact `f32`, `K * 4` bytes
+- **strict lower triangle** — `bf16`, mirrored on read, index `i*(i-1)/2 + j` for `i > j`
+
+`bf16` conversion is **truncation** (`f32_bits >> 16`), not round-to-nearest.
+
+These are full Hessians (`XX^T`), not diagonals. A 496-tensor 27B package is
+~31.5 GB, averaging ~63 MB per Hessian; a diagonal-only package would be ~10 MB
+total.
+
+## REQUIRED: project to PSD before Cholesky
+
+**Reading a stored Hessian and running Cholesky directly will fail on roughly
+half of all tensors.** This is measured, not theoretical.
+
+`bf16` off-diagonals perturb the Gram matrix out of positive-semidefiniteness by
+an amount comparable to `mean(diag(H))`. Measured on
+`layers.1.linear_attn.in_proj_qkv` (K=5120) of the Qwen3.8-27B package:
+
+| quantity | value |
+|---|---|
+| `mean(diag(H))` | 9.951136e-01 |
+| off-diagonal RMS | 7.889832e-01 |
+| Cauchy-Schwarz violations | **0 / 1,047,552** |
+| leading 1024x1024 `lambda_min` | **-2.088e-01** (353 of 1024 eigenvalues negative) |
+| negative eigenvalue energy | **0.99%** of positive |
+
+Zero Cauchy-Schwarz violations means the data is structurally sound — this is
+precision noise, not corruption or a normalization error.
+
+Because `lambda_min` scales roughly as `sqrt(K)`, failure is K-dependent:
+
+| K | `-lambda_min / mean(diag)` | observed GPTQ outcome |
+|---|---|---|
+| 5120 | ~1.0 | ~50% of tensors fail Cholesky |
+| 6144 | ~1.1 | passes only at near-max damping, correction collapses to ~0 |
+| 17408 | ~1.8 | fails always |
+
+Raising the damping cap is **not** a fix: once `damp ~ mean(diag(H))`,
+`H + damp*I` is diagonal-dominated and GPTQ degenerates to round-to-nearest.
+
+### The fix
+
+Eigendecompose, clip negative eigenvalues to zero, rebuild:
+
+```python
+evals, evecs = numpy.linalg.eigh(H)
+H = (evecs * numpy.clip(evals, 0.0, None)) @ evecs.T
+H = (H + H.T) / 2
+```
+
+Measured effect on the tensor above:
+
+| | Cholesky succeeds at damp |
+|---|---|
+| raw stored | 1.0065 (exceeds a `1.0 * mean(diag)` cap, so it fails) |
+| PSD-projected | **0.01** |
+
+The projection changes the matrix by **0.1004%** (relative Frobenius), which is
+*below* `bf16`'s own ~0.391% quantization step. It removes less than the storage
+format already discarded: denoising, not signal loss. Cost is one `O(K^3)`
+eigendecomposition per tensor — the same order as the Cholesky it enables.
+
+## `.ref_*.bin` — KLD reference tensors
+
+Fixed teacher logits for paired KLD/PPL scoring. **top-k 256, not
+full-vocabulary**, so absolute KLD values are *not* comparable to
+`llama.cpp --kl-divergence` tables. Relative ranking within this instrument is
+valid; absolute cross-table comparison is not. Oracle NLL/PPL were computed from
+full-vocab teacher logits before top-k serialization.
+
+Protocol: `n_ctx` 2048, 24 chunks, prefill scoring mode.
+
+| suffix | corpus |
+|---|---|
+| `ref_wt2` | WikiText-2 — prose. An out-of-distribution tripwire. |
+| `ref_v6sel` | Held-out rendered chat-template conversations, 0% prose. |
+| `ref_ag` | AG News. |
+
+**A prose reference alone is insufficient for a chat model.** It compresses arm
+margins by roughly 3x relative and can invert a PPL ranking. Score against a
+deployment-distribution reference alongside the prose tripwire.