The two files represent the same model architecture, with a broad, dense update spread across 3 model region(s) — 3 transformer block(s).
| Classification | Meaning |
|---|---|
| IDENTICAL | Files are byte-for-byte identical (same SHA-256) — includes both weights and metadata. |
| METADATA_ONLY | All tensor/weight values are identical; only metadata or padding bytes differ. |
| LOCALIZED_WEIGHT_EDIT | Few tensors/regions changed — consistent with a targeted edit. |
| FINE_TUNED | Distributed changes across many tensors, same architecture. |
| ARCHITECTURE_MODIFIED | Tensors, shapes, or regions added/removed/changed. |
| INCOMPATIBLE | Too few shared tensors for a meaningful comparison. |
"Weight values differing" counts individual values where A ≠ B, not tensors — a tensor with 1% of its values changed still counts as "1 tensor changed" above, but contributes proportionally here. This counts positions that differ, not which values a training/optimizer process selected to update. Quantized tensors (Q4_K, IQ-series, etc.) are counted as fully changed in this total, since we can't isolate which individual values differ without dequantizing block-quantized formats.
Each cell is one transformer block (or top-level tensor group). Orange = at least one tensor in that block differs.
| Layer | Changed / Total |
|---|---|
| blk.0 | 3 / 7 |
| blk.1 | 3 / 7 |
| blk.2 | 3 / 7 |
| Tensor | Kind | Shape | L2 diff | Changed elements | Max / mean abs diff |
|---|---|---|---|---|---|
| blk.1.ffn_down.weight | VALUE_DIFF | [16, 8] | 48.35 (461.3877% rel.) | 128 / 128 (100.0%) | 11.03 / 3.569 |
More stats (percentiles, cosine similarity, distribution)RMS diff
4.273 p50 abs diff
3.401 p95 abs diff
7.751 p99 abs diff
9.346 Cosine similarity
-0.085137 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
30
25
20
28
12
9
2
2
≤1.4 ≤2.8 ≤4.2 ≤5.6 ≤6.9 ≤8.3 ≤9.7 ≤11
|
|||||
| blk.0.ffn_up.weight | VALUE_DIFF | [8, 16] | 49.04 (459.9096% rel.) | 128 / 128 (100.0%) | 11.09 / 3.462 |
More stats (percentiles, cosine similarity, distribution)RMS diff
4.334 p50 abs diff
2.79 p95 abs diff
8.736 p99 abs diff
10.48 Cosine similarity
0.126821 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
32
32
20
18
12
5
7
2
≤1.4 ≤2.8 ≤4.2 ≤5.6 ≤7 ≤8.3 ≤9.7 ≤11
|
|||||
| blk.1.ffn_up.weight | VALUE_DIFF | [8, 16] | 50.38 (436.3675% rel.) | 128 / 128 (100.0%) | 15.81 / 3.668 |
More stats (percentiles, cosine similarity, distribution)RMS diff
4.453 p50 abs diff
3.207 p95 abs diff
8.244 p99 abs diff
11.37 Cosine similarity
-0.104542 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
35
40
34
11
5
2
0
1
≤2 ≤4 ≤5.9 ≤7.9 ≤9.9 ≤12 ≤14 ≤16
|
|||||
| blk.2.ffn_up.weight | VALUE_DIFF | [8, 16] | 50.15 (425.8809% rel.) | 128 / 128 (100.0%) | 12.17 / 3.567 |
More stats (percentiles, cosine similarity, distribution)RMS diff
4.433 p50 abs diff
3.306 p95 abs diff
7.994 p99 abs diff
11.67 Cosine similarity
0.000872 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
38
20
26
24
11
6
0
3
≤1.6 ≤3.1 ≤4.6 ≤6.1 ≤7.6 ≤9.1 ≤11 ≤12
|
|||||
| blk.1.ffn_gate.weight | VALUE_DIFF | [8, 16] | 46.43 (393.8017% rel.) | 128 / 128 (100.0%) | 11.88 / 3.322 |
More stats (percentiles, cosine similarity, distribution)RMS diff
4.104 p50 abs diff
3.088 p95 abs diff
8.105 p99 abs diff
10.75 Cosine similarity
0.088717 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
34
29
31
18
6
6
2
2
≤1.5 ≤3 ≤4.5 ≤5.9 ≤7.4 ≤8.9 ≤10 ≤12
|
|||||
| blk.0.ffn_down.weight | VALUE_DIFF | [16, 8] | 44.65 (385.572% rel.) | 128 / 128 (100.0%) | 12.63 / 3.263 |
More stats (percentiles, cosine similarity, distribution)RMS diff
3.947 p50 abs diff
2.827 p95 abs diff
7.541 p99 abs diff
9.19 Cosine similarity
0.102216 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
33
38
30
14
8
4
0
1
≤1.6 ≤3.2 ≤4.8 ≤6.3 ≤7.9 ≤9.5 ≤11 ≤13
|
|||||
| blk.0.ffn_gate.weight | VALUE_DIFF | [8, 16] | 45.19 (383.2007% rel.) | 128 / 128 (100.0%) | 9.151 / 3.309 |
More stats (percentiles, cosine similarity, distribution)RMS diff
3.995 p50 abs diff
2.973 p95 abs diff
7.349 p99 abs diff
8.627 Cosine similarity
0.093739 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
23
25
23
22
13
11
7
4
≤1.2 ≤2.3 ≤3.4 ≤4.6 ≤5.7 ≤6.9 ≤8 ≤9.2
|
|||||
| blk.2.ffn_down.weight | VALUE_DIFF | [16, 8] | 46.42 (373.8425% rel.) | 128 / 128 (100.0%) | 12.03 / 3.281 |
More stats (percentiles, cosine similarity, distribution)RMS diff
4.103 p50 abs diff
2.751 p95 abs diff
8.593 p99 abs diff
9.353 Cosine similarity
0.057604 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
36
32
23
20
7
6
3
1
≤1.5 ≤3 ≤4.5 ≤6 ≤7.5 ≤9 ≤11 ≤12
|
|||||
| blk.2.ffn_gate.weight | VALUE_DIFF | [8, 16] | 42.47 (346.5571% rel.) | 128 / 128 (100.0%) | 10.36 / 3.022 |
More stats (percentiles, cosine similarity, distribution)RMS diff
3.754 p50 abs diff
2.518 p95 abs diff
6.866 p99 abs diff
9.74 Cosine similarity
0.048122 Distribution of |diff| across all 128 elements (8 buckets, upper bound shown) — exact counts above each bar:
36
28
22
20
11
7
1
3
≤1.3 ≤2.6 ≤3.9 ≤5.2 ≤6.5 ≤7.8 ≤9.1 ≤10
|
|||||
For float tensors (F32/F16/BF16/F64, dequantized to float32 first): L2 diff = ‖B − A‖₂ over all elements; relative % = that divided by ‖A‖₂ (the reference tensor's own norm), ×100; RMS diff (in "More stats") = L2 diff / √N. "Changed elements" counts positions where A ≠ B exactly — a high L2 with few changed elements means a few large changes; a lower L2 with many changed elements means many small ones, which "L2 diff" alone can't distinguish. Rows below are ranked by relative L2, not absolute L2 — absolute L2 scales with tensor size (√N), so a 38M-element embedding with tiny per-element noise would otherwise rank above a 2M-element tensor with a much larger relative change, which is what actually matters for spotting the most significantly changed tensor. Quantized tensors (Q4_K, IQ-series, etc.) show a byte-level diff % instead — a true numeric norm on raw quantized bytes isn't meaningful, so it's reported separately and isn't comparable to the L2 numbers above.
{
"block_map": [
{
"changed": true,
"layer": "blk.0"
},
{
"changed": true,
"layer": "blk.1"
},
{
"changed": true,
"layer": "blk.2"
},
{
"changed": false,
"layer": "output_norm.weight"
},
{
"changed": false,
"layer": "token_embd.weight"
}
],
"compat_note": "Models structurally identical: 23 shared tensors.",
"compatible": true,
"differences": [
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": -0.085137,
"diff_histogram": [
{
"count": 30,
"range_max": 1.44324686
},
{
"count": 25,
"range_max": 2.81244523
},
{
"count": 20,
"range_max": 4.1816436
},
{
"count": 28,
"range_max": 5.55084197
},
{
"count": 12,
"range_max": 6.92004034
},
{
"count": 9,
"range_max": 8.28923871
},
{
"count": 2,
"range_max": 9.65843708
},
{
"count": 2,
"range_max": 11.02763546
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 48.34777960028881,
"l2_pct": 461.3877,
"magnitude": 461.3877,
"max_abs_diff": 11.02763545513153,
"mean_abs_diff": 3.56884781,
"method": "l2_float",
"name": "blk.1.ffn_down.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 3.40100721,
"p95_abs_diff": 7.75083669,
"p99_abs_diff": 9.34553879,
"rms_diff": 4.27338035,
"shape": [
16,
8
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.126821,
"diff_histogram": [
{
"count": 32,
"range_max": 1.44112305
},
{
"count": 32,
"range_max": 2.81928994
},
{
"count": 20,
"range_max": 4.19745683
},
{
"count": 18,
"range_max": 5.57562372
},
{
"count": 12,
"range_max": 6.95379061
},
{
"count": 5,
"range_max": 8.3319575
},
{
"count": 7,
"range_max": 9.7101244
},
{
"count": 2,
"range_max": 11.08829129
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 49.038621832821356,
"l2_pct": 459.9096,
"magnitude": 459.9096,
"max_abs_diff": 11.08829128742218,
"mean_abs_diff": 3.46229797,
"method": "l2_float",
"name": "blk.0.ffn_up.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 2.78999221,
"p95_abs_diff": 8.73577198,
"p99_abs_diff": 10.48407966,
"rms_diff": 4.33444275,
"shape": [
8,
16
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": -0.104542,
"diff_histogram": [
{
"count": 35,
"range_max": 1.99164162
},
{
"count": 40,
"range_max": 3.96637516
},
{
"count": 34,
"range_max": 5.94110871
},
{
"count": 11,
"range_max": 7.91584225
},
{
"count": 5,
"range_max": 9.89057579
},
{
"count": 2,
"range_max": 11.86530934
},
{
"count": 0,
"range_max": 13.84004288
},
{
"count": 1,
"range_max": 15.81477642
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 50.38441102910399,
"l2_pct": 436.3675,
"magnitude": 436.3675,
"max_abs_diff": 15.814776420593262,
"mean_abs_diff": 3.66773413,
"method": "l2_float",
"name": "blk.1.ffn_up.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 3.20666713,
"p95_abs_diff": 8.24359455,
"p99_abs_diff": 11.36870443,
"rms_diff": 4.45339484,
"shape": [
8,
16
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.000872,
"diff_histogram": [
{
"count": 38,
"range_max": 1.55478225
},
{
"count": 20,
"range_max": 3.07109179
},
{
"count": 26,
"range_max": 4.58740132
},
{
"count": 24,
"range_max": 6.10371086
},
{
"count": 11,
"range_max": 7.6200204
},
{
"count": 6,
"range_max": 9.13632993
},
{
"count": 0,
"range_max": 10.65263947
},
{
"count": 3,
"range_max": 12.16894901
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 50.15446937195353,
"l2_pct": 425.8809,
"magnitude": 425.8809,
"max_abs_diff": 12.168949007987976,
"mean_abs_diff": 3.56735934,
"method": "l2_float",
"name": "blk.2.ffn_up.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 3.30644929,
"p95_abs_diff": 7.99442142,
"p99_abs_diff": 11.6722097,
"rms_diff": 4.43307067,
"shape": [
8,
16
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.088717,
"diff_histogram": [
{
"count": 34,
"range_max": 1.50123033
},
{
"count": 29,
"range_max": 2.98328514
},
{
"count": 31,
"range_max": 4.46533995
},
{
"count": 18,
"range_max": 5.94739475
},
{
"count": 6,
"range_max": 7.42944956
},
{
"count": 6,
"range_max": 8.91150436
},
{
"count": 2,
"range_max": 10.39355917
},
{
"count": 2,
"range_max": 11.87561397
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 46.43483515742366,
"l2_pct": 393.8017,
"magnitude": 393.8017,
"max_abs_diff": 11.87561397254467,
"mean_abs_diff": 3.32163602,
"method": "l2_float",
"name": "blk.1.ffn_gate.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 3.08768418,
"p95_abs_diff": 8.10512584,
"p99_abs_diff": 10.7502869,
"rms_diff": 4.10429835,
"shape": [
8,
16
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.102216,
"diff_histogram": [
{
"count": 33,
"range_max": 1.63601811
},
{
"count": 38,
"range_max": 3.20672694
},
{
"count": 30,
"range_max": 4.77743576
},
{
"count": 14,
"range_max": 6.34814459
},
{
"count": 8,
"range_max": 7.91885342
},
{
"count": 4,
"range_max": 9.48956224
},
{
"count": 0,
"range_max": 11.06027107
},
{
"count": 1,
"range_max": 12.6309799
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 44.650600247563105,
"l2_pct": 385.572,
"magnitude": 385.572,
"max_abs_diff": 12.630979895591736,
"mean_abs_diff": 3.26294409,
"method": "l2_float",
"name": "blk.0.ffn_down.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 2.82698542,
"p95_abs_diff": 7.5414599,
"p99_abs_diff": 9.18996341,
"rms_diff": 3.94659278,
"shape": [
16,
8
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.093739,
"diff_histogram": [
{
"count": 23,
"range_max": 1.15799958
},
{
"count": 25,
"range_max": 2.29985607
},
{
"count": 23,
"range_max": 3.44171256
},
{
"count": 22,
"range_max": 4.58356905
},
{
"count": 13,
"range_max": 5.72542554
},
{
"count": 11,
"range_max": 6.86728203
},
{
"count": 7,
"range_max": 8.00913852
},
{
"count": 4,
"range_max": 9.15099502
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 45.19436449791195,
"l2_pct": 383.2007,
"magnitude": 383.2007,
"max_abs_diff": 9.150995016098022,
"mean_abs_diff": 3.30926148,
"method": "l2_float",
"name": "blk.0.ffn_gate.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 2.97331089,
"p95_abs_diff": 7.34850987,
"p99_abs_diff": 8.62692828,
"rms_diff": 3.9946552,
"shape": [
8,
16
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.057604,
"diff_histogram": [
{
"count": 36,
"range_max": 1.50985327
},
{
"count": 32,
"range_max": 3.01310055
},
{
"count": 23,
"range_max": 4.51634783
},
{
"count": 20,
"range_max": 6.01959512
},
{
"count": 7,
"range_max": 7.5228424
},
{
"count": 6,
"range_max": 9.02608968
},
{
"count": 3,
"range_max": 10.52933697
},
{
"count": 1,
"range_max": 12.03258425
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 46.417598365364874,
"l2_pct": 373.8425,
"magnitude": 373.8425,
"max_abs_diff": 12.032584249973297,
"mean_abs_diff": 3.28121227,
"method": "l2_float",
"name": "blk.2.ffn_down.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 2.75140455,
"p95_abs_diff": 8.59339845,
"p99_abs_diff": 9.35317592,
"rms_diff": 4.10277482,
"shape": [
16,
8
],
"total_elements": 128
},
{
"changed_elements": 128,
"changed_elements_pct": 100.0,
"cosine_similarity": 0.048122,
"diff_histogram": [
{
"count": 36,
"range_max": 1.29524835
},
{
"count": 28,
"range_max": 2.59036128
},
{
"count": 22,
"range_max": 3.88547421
},
{
"count": 20,
"range_max": 5.18058714
},
{
"count": 11,
"range_max": 6.47570007
},
{
"count": 7,
"range_max": 7.770813
},
{
"count": 1,
"range_max": 9.06592593
},
{
"count": 3,
"range_max": 10.36103886
}
],
"dtype": "F32",
"kind": "VALUE_DIFF",
"l2_diff": 42.473712613471776,
"l2_pct": 346.5571,
"magnitude": 346.5571,
"max_abs_diff": 10.361038863658905,
"mean_abs_diff": 3.02209579,
"method": "l2_float",
"name": "blk.2.ffn_gate.weight",
"nan_or_inf_introduced": false,
"p50_abs_diff": 2.51829856,
"p95_abs_diff": 6.86554278,
"p99_abs_diff": 9.74049213,
"rms_diff": 3.75418128,
"shape": [
8,
16
],
"total_elements": 128
}
],
"file_a": {
"filename": "overfit-epoch1.gguf",
"sha256": "d1a47e6c104c015b241cdfe97e8f754ac85078d3d8e1f0d0827e2222e709c0be",
"size_bytes": 11264,
"tensor_count": 23
},
"file_b": {
"filename": "overfit-epoch10.gguf",
"sha256": "cabe29e8da61b73b795650dce4468f0ee169ea82b4146ba3389090c0abfef042",
"size_bytes": 11264,
"tensor_count": 23
},
"global_summary": {
"max_l2_tensor_name": "blk.1.ffn_down.weight",
"max_tensor_l2_pct": 461.3877,
"median_cosine_similarity": 0.057604,
"median_tensor_l2_pct": 393.8017,
"nan_or_inf_introduced": false,
"tensors_changed": 9,
"total_weight_values": 2184,
"weight_values_differing": 1152,
"weight_values_differing_pct": 52.7473
},
"layer_summary": [
{
"layer": "blk.0",
"tensors_changed": 3,
"total_tensors": 7
},
{
"layer": "blk.1",
"tensors_changed": 3,
"total_tensors": 7
},
{
"layer": "blk.2",
"tensors_changed": 3,
"total_tensors": 7
}
],
"missing_in_a": [],
"missing_in_b": [],
"summary": {
"different": 9,
"identical": 14,
"missing_in_a": 0,
"missing_in_b": 0,
"total_common": 23
},
"verdict": {
"changed_area": "3 model region(s) \u2014 3 transformer block(s)",
"classification": "FINE_TUNED",
"explanation": "9 of 23 tensor(s) (39.1%) differ in value, spread across 3 model region(s) \u2014 3 transformer block(s). The architecture is unchanged, but the change is too broad to call a single localized edit \u2014 consistent with a general fine-tune. The breadth/distribution of the change is directly measured (high confidence); the training procedure that produced it cannot be proven from artifact differences alone (medium confidence).",
"likely_origin": "General fine-tuning (broad weight update, same architecture)",
"likely_origin_confidence": "Medium",
"narrative_summary": "The two files represent the same model architecture, with a broad, dense update spread across 3 model region(s) \u2014 3 transformer block(s).",
"observed_pattern": "Broad distributed weight update \u2014 9 of 23 tensor(s) (39.1%) changed across 3 model region(s) \u2014 3 transformer block(s)",
"observed_pattern_confidence": "High",
"structural_impact": "None detected"
}
}