Update public v1 dataset with validated M4 Qwen and M4 CPU reruns

README.md

@@ -42,7 +42,7 @@ Controlled inference benchmark dataset for **7 GGUF K-quant quantization variant
 | Apple M4 Mac | Apple M4 (ARM, 10-core) | 16 GB unified | llama.cpp Metal |
 | HP Pavilion x86 | Intel Core i5-1235U (12th gen) | 16 GB DDR4 | llama.cpp CPU |
 
-**4,407 total records** across 5 splits. All inference records are non-warmup,
+**3,395 total records** across 5 splits. All published inference records are non-warmup,
 success-status runs collected under controlled thermal conditions. Contaminated
 and failed records are archived separately and not included here.
 
@@ -71,7 +71,7 @@ and failed records are archived separately and not included here.
 
 ## Splits
 
-### `pixel_inference` — 2,875 rows
+### `pixel_inference` — 1,819 rows
 Pixel 6a (ARM, CPU backend) inference runs.
 
 | Column | Type | Description |
@@ -84,10 +84,13 @@ Pixel 6a (ARM, CPU backend) inference runs.
 | `trial` | int | Trial index within the experiment |
 | `threads` | int | CPU thread count (null = default 4) |
 | `decode_tps` | float | Decode throughput (tokens/second) |
+| `decode_tps_std` | float | Decode TPS standard deviation for pre-aggregated rows; null for individual-trial rows |
 | `prefill_tps` | float | Prefill throughput (tokens/second) |
+| `prefill_tps_std` | float | Prefill TPS standard deviation for pre-aggregated rows; null for individual-trial rows |
 | `ttft_s` | float | Time to first token (seconds) — populated for standard_sweep only |
 | `e2e_s` | float | End-to-end latency (seconds) — populated for standard_sweep only |
 | `n_output_tokens` | int | Number of generated tokens |
+| `n_trials` | int | Number of trials represented by the row (`1` for individual trial rows; `5`/`10` for aggregated rows) |
 | `experiment_type` | string | `cliff_sweep` \| `standard_sweep` \| `thread_sweep` \| `kv_cache_quant` |
 | `kv_quant` | string | KV cache quantization type (`null` = default, `"q8_0"` = quantized) |
 | `ngl` | int | GPU layers (null for CPU runs) |
@@ -96,28 +99,34 @@ Pixel 6a (ARM, CPU backend) inference runs.
 
 **experiment_type values:**
 - `cliff_sweep` — context length varied to characterise KV-cache collapse (canonical n=10)
-- `standard_sweep` — fixed 4 context windows (256/512/1024/2048), 13 trials, 2 warmup
+- `standard_sweep` — fixed 4 context windows (256/512/1024/2048), canonical TPS sweep
 - `thread_sweep` — Q4\_K\_M at threads=1/2/4/8, ctx=256, 15 trials
 - `kv_cache_quant` — KV cache set to q8\_0 to test collapse mitigation
 
 ---
 
-### `m4_inference` — 1,021 rows
-Apple M4 Mac inference runs. Contains two backend configurations:
+### `m4_inference` — 1,035 rows
+Apple M4 Mac inference runs. Contains Metal GPU and CPU backend configurations:
 
-- **Metal GPU** (931 rows) — `backend = "Metal"`, `ngl = 99`. Includes Llama 3.2 3B and Qwen 2.5 1.5B.
+- **Metal GPU, Llama 3.2 3B** (840 rows) — `backend = "Metal"`, `ngl = 99`.
   Cliff sweep covers ctx=1024–2048 (13 points, n=5 trials). Results: flat profile on Metal
   (all variants within ±9%), confirming no KV-cache cliff on GPU-accelerated inference.
-- **CPU** (90 rows) — `backend = "CPU"`, `ngl = 0`, `threads = 4`. Llama 3.2 3B only.
+- **Metal GPU, Qwen 2.5 1.5B** (98 rows) — promoted clean extension runs:
+  - **TPS sweep** (7 rows, `standard_sweep`, `context_len = 0`): tg128 decode, n=10,
+    `results/m4_qwen_tps_20260415_130955/`. Q2\_K=36.56 tok/s, Q8\_0=21.50 tok/s.
+  - **Cliff sweep** (91 rows, `cliff_sweep`): ctx=1024–2048, 13 contexts × 7 variants,
+    n=5, `results/m4_qwen_cliff_20260416_021323/`. Q2\_K changes −32.7% from
+    ctx=1024→2048; Q8\_0 is flat (+1.0%).
+- **CPU, Llama 3.2 3B** (97 rows) — `backend = "CPU"`, `ngl = 0`, `threads = 4`.
   - **Cliff sweep** (88 rows): ctx=256–2048 (13 points, pre-aggregated n\_trials=5 per ctx).
     Collected 2026-04-09. 3 outlier points excluded (Q5\_K\_M ctx=2048 OOM, Q6\_K ctx=1536
     CV=81%, Q8\_0 ctx=2048 CV=99%). Results: significant context-dependent degradation
     on M4 CPU (Q2\_K −13%, Q3\_K\_M −54%, Q4\_K\_S −53%, Q6\_K −60% from ctx=256→2048).
     Note: ctx=256 cliff baseline may be inflated by CPU boost state at start of each variant's sweep.
   - **TPS sweep** (7 rows, `experiment_type = "standard_sweep"`, `context_len = 0`): pure decode
-    reference (n\_prompt=0, n\_gen=128, n=10 trials, 2026-04-06). Thermally settled baseline.
-    Throughput ordering: Q4\_K\_S (13.16) > Q8\_0 (12.60) > Q4\_K\_M (12.51) > Q2\_K (12.31)
-    > Q3\_K\_M (11.48) > Q5\_K\_M (10.59) > Q6\_K (9.29) tok/s. Non-monotonic: Metal reversal
+    reference (n\_prompt=0, n\_gen=128, n=10 trials, 2026-04-15 clean rerun).
+    Throughput ordering: Q2\_K (27.08) > Q4\_K\_S (25.68) > Q3\_K\_M (23.67) > Q4\_K\_M (22.29)
+    > Q8\_0 (18.80) > Q5\_K\_M (15.80) > Q6\_K (14.97) tok/s. Non-monotonic: Metal reversal
     (Q4\_K\_S fastest) confirmed on M4 CPU as well; Q6\_K remains slowest.
 
 Same columns as `pixel_inference`.
@@ -143,14 +152,15 @@ Same columns as `pixel_inference`. `backend = "CPU"`, `threads = 6`.
 
 ---
 
-### `quality_benchmarks` — 105 rows
-Accuracy scores on 6 NLP benchmarks for 7 quantization variants on Pixel 6a.
+### `quality_benchmarks` — 128 rows
+Accuracy scores on 6 NLP benchmarks for 7 quantization variants across Pixel 6a and x86 i5-1235U,
+including both standard and imatrix-calibrated variants.
 
 | Column | Type | Description |
 |---|---|---|
-| `benchmark` | string | `arc_challenge` \| `arc_easy` \| `boolq` \| `hellaswag` \| `mmlu` \| `truthfulqa` \| `custom_qa` |
+| `benchmark` | string | `arc_challenge` \| `arc_easy` \| `boolq` \| `hellaswag` \| `mmlu` \| `truthfulqa` |
 | `variant` | string | GGUF quantization variant |
-| `device` | string | `"Pixel6a"` |
+| `device` | string | `"Pixel6a"` or `"x86"` |
 | `model` | string | Model name |
 | `calibration` | string | `"standard"` or `"imatrix"` (importance-weighted) |
 | `accuracy_pct` | float | Accuracy percentage (0–100) |
@@ -158,30 +168,33 @@ Accuracy scores on 6 NLP benchmarks for 7 quantization variants on Pixel 6a.
 | `total` | int | Total questions evaluated |
 | `status` | string | `"success"` for all included rows |
 
+**Device coverage:**
+- `Pixel6a`: 44 standard rows + 42 imatrix rows (ARC-Easy imatrix excluded — known parser artifact producing 100% for all variants)
+- `x86`: all 6 benchmarks × 7 variants (standard only; no imatrix)
+
 **Benchmark sample sizes:** 100 questions each (random sample from official test sets).
 BoolQ imatrix calibration covers all 7 variants. TruthfulQA imatrix data collected for
-Q2\_K and Q3\_K\_M only.
+all 7 variants.
 
 ---
 
-### `perplexity` — 7 rows
-WikiText-2 perplexity scores for Llama 3.2 3B Instruct on Pixel 6a.
+### `perplexity` — 14 rows
+WikiText-2 perplexity scores for Llama 3.2 3B Instruct. Covers both Pixel 6a and x86 i5-1235U measurements.
 
 | Column | Type | Description |
 |---|---|---|
 | `variant` | string | GGUF quantization variant |
 | `model` | string | Model name |
-| `device` | string | `"Pixel6a"` |
-| `perplexity` | float | WikiText-2 perplexity (lower = better); null if not evaluated |
+| `device` | string | `"Pixel6a"` or `"x86"` |
+| `perplexity` | float | WikiText-2 perplexity (lower = better) |
 | `perplexity_status` | string | `"success"` or `"not_evaluated"` |
-| `corpus` | string | `"wikitext2_full"` (~285K tokens) or `"wikitext2_sample"` (~12K tokens) |
+| `corpus` | string | `"wikitext2_full"` (~290K tokens) or `"wikitext2_sample"` (~12K tokens) |
 | `tokens_approx` | int | Approximate token count used |
-| `note` | string | Reason if not\_evaluated |
+| `note` | string | Measurement notes |
 
-> **Important:** Q2\_K and Q3\_K\_M were evaluated on the full WikiText-2 corpus;
-> Q4\_K\_M, Q6\_K, Q8\_0 on a 12K-token sample. Do not directly compare perplexity
-> values across these two groups without accounting for corpus size effects.
-> Q4\_K\_S and Q5\_K\_M were added after the initial sweep and are marked `not_evaluated`.
+> **All 7 variants have full-corpus PPL values.** Q2\_K and Q3\_K\_M measured on Pixel 6a (full corpus, ~285K tokens).
+> Q4\_K\_S, Q4\_K\_M, Q5\_K\_M, Q6\_K, Q8\_0 measured on x86 i5-1235U (full corpus, ~290K tokens).
+> Pixel 6a also has sample-corpus (~12K tokens) measurements for Q4\_K\_M, Q6\_K, Q8\_0 (retained for reference).
 
 ---
 
@@ -267,12 +280,13 @@ print(threads.groupby("threads")["decode_tps"].agg(["mean", "std"]))
 1. **Pixel 6a primary focus** — x86 and M4 coverage is less comprehensive than Pixel;
    x86 has n=5 trials for cliff sweep but no thread sweep, no kv_cache_quant experiments
 2. **x86 Qwen limited to standard_sweep** — Qwen 2.5 1.5B on x86 provides decode TPS reference
-   at ctx=256 only; no cliff sweep, thread sweep, or quality data for Qwen on x86
+   at ctx=256 only. Two cliff reruns were attempted and pushed, but are excluded because the
+   result files contain missing/zero-throughput rows at larger contexts
 3. **Perplexity corpus inconsistency** — see note in perplexity split above
 4. **No power/energy data** — `/proc` interfaces on Pixel 6a are unreliable without root;
    battery drain proxy metrics were collected but not included in this release
-5. **Single model family for quality benchmarks** — quality data (BoolQ, HellaSwag, etc.)
-   collected on Pixel 6a only; no cross-device quality comparison
+5. **Single model family for quality benchmarks** — quality data covers Llama 3.2 3B only;
+   Pixel 6a and x86 rows are included, but there is no validated M4 or Qwen quality split
 6. **llama.cpp version** — builds used llama.cpp circa February–April 2026;
    results may differ with significantly newer versions
 

m4_inference.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10ee4abc859e12e8ee543ba0e4baf206ba040ebee1b42eb6d59ebee8ae874b8a
-size 24632
+oid sha256:2a89fd64badf23fd5d2d5ac7452b47049a0d67ae234c6776517e208513fb899b
+size 32843

perplexity.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b60d15ef974ed5234c7de5b8d56bbc382e970c96e4657a903700981684de93a0
-size 5590
+oid sha256:9ef88b45a6deffd2ea0d0b760036be9f6ca1498f1d2654a783e413bedc7fe562
+size 5689

pixel_inference.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3e3a41a84a41a910bc0717ee428167d8265447252dff1b30fb20302765e1b897
-size 53995
+oid sha256:085c4ed285ff2aff44287af154807a8b654532f4109d1088b24977236d368e80
+size 35253

quality_benchmarks.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b9b977de9d6034911446280e4f3a571a8132b8111a126d09917ee64a07235d4
-size 6404
+oid sha256:4b212aa74f65084e5031b95fe15d3f12b508c2dc89a6eb3a9ae2084e360e7a29
+size 6371

x86_inference.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:68a285d4efb3c7b7d3d9b235d7a116ba8153bfc893428be37a9897ef844ab200
-size 16530
+oid sha256:fb319c6c76e8ec7c6b9e7fbf0a4be9359072df46b6c1ec6deed2decf5692814e
+size 18330