Update README and 1 associated file(s)

README.md

@@ -520,800 +520,5 @@ configs:
   - split: train
     path: shard_6/train-*
 ---
-
-# Encoder/decoder trial: encoder-marginal report
-
-- Dataset: `G-reen/encoder-decoder-trial-stat`
-- Rows analysed: 18,255 (every kept (encoder, decoder, source row) triple; source `G-reen/cc-re-2021-filtered` shard 0, 2000 rows of at most 4000 words)
-- Prompt file: `prompts/indirect_reference_dataset_train.json` (turn 0 encodes the document, turn 1 reconstructs it from the encoding alone)
-- Encoders: 9 (granite-4.2-30b-nvfp4 [0], Ornith-1.5-35B-A3B-NVFP4 [1], Llama-3.3-70B-Instruct-NVFP4 [2], Qwen3.8-27B-AWQ-INT4 [3], Mistral-Small-4-119B-2603-NVFP4 [4], gemma-4-31B-it-AWQ-4bit [5], Laguna-S-2.1-NVFP4 [6], claude-sonnet-5 [8], gpt-5.6-terra [9])
-- Decoders: 7 (granite-4.2-30b-nvfp4 [0], Ornith-1.5-35B-A3B-NVFP4 [1], Llama-3.3-70B-Instruct-NVFP4 [2], Qwen3.8-27B-AWQ-INT4 [3], Mistral-Small-4-119B-2603-NVFP4 [4], gemma-4-31B-it-AWQ-4bit [5], Laguna-S-2.1-NVFP4 [6])
-
-Models (trial index: config):
-- 0: `config/gen/train/shard_0.toml`
-- 1: `config/gen/train/shard_1.toml`
-- 2: `config/gen/train/shard_2.toml`
-- 3: `config/gen/train/shard_3.toml`
-- 4: `config/gen/train/shard_4.toml`
-- 5: `config/gen/train/shard_5.toml`
-- 6: `config/gen/train/shard_6.toml`
-- 7: `config/gen/train/shard_7.toml` (decode only)
-- 8: `config/encdec/claude_sonnet_5.toml` (encode only)
-- 9: `config/encdec/gpt_5_6_terra.toml` (encode only)
-
-Statistics:
-- every configured statistic was present.
-
-## Summary: per-encoder means, marginalised over decoders
-
-Every encoder encoded the same 2000 rows and every decoder decoded the first 300 of every encoder's encodings (the same source rows for every encoder), so each encoder's row is an average over the same decoders and source texts. Values are the decoder-balanced means (mean of the per-decoder means); `Rows` counts the kept rows behind each.
-For reference, the human source texts score 0.0616 (std 0.0942) on the same EditLens model.
-
-| Encoder | Editlens Score | Editlens Bucket | Cosdist | Jaccard 1 | Levenshtein | Softngram | Bertscore | Moverscore | Reranker | Rows |
-|---|---|---|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.6851 | 2.0366 | 0.2720 | 0.7668 | 2953.7796 | 0.7875 | 0.1751 | 0.6318 | -2.0271 | 2,017 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6311 | 1.8620 | 0.2825 | 0.7186 | 2830.1266 | 0.7293 | 0.1639 | 0.6030 | -1.8051 | 2,023 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8029 | 2.4331 | 0.3843 | 0.8341 | 3043.3284 | 0.8686 | 0.2013 | 0.6854 | 1.7986 | 2,013 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.7214 | 2.1729 | 0.2912 | 0.7946 | 3020.2663 | 0.8372 | 0.1822 | 0.6507 | -1.6815 | 2,048 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.7510 | 2.2569 | 0.3192 | 0.8045 | 3198.6141 | 0.8353 | 0.1872 | 0.6591 | -0.4909 | 2,020 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.7610 | 2.2775 | 0.3312 | 0.8345 | 3053.2435 | 0.8940 | 0.1955 | 0.6787 | -0.6185 | 2,043 |
-| Laguna-S-2.1-NVFP4 [6] | 0.7304 | 2.1809 | 0.3352 | 0.8020 | 3184.1194 | 0.8390 | 0.1896 | 0.6629 | -0.4826 | 2,014 |
-| ❗ claude-sonnet-5 [8] | 0.6172 | 1.7905 | 0.2419 | 0.7545 | 3155.4410 | 0.7796 | 0.1678 | 0.6204 | -2.7809 | 2,038 |
-| gpt-5.6-terra [9] | 0.6196 | 1.8320 | 0.2224 | 0.7442 | 3330.6634 | 0.7693 | 0.1637 | 0.6117 | -3.4548 | 2,039 |
-
-✔️ marks the highest EditLens score (most AI-like reconstructions), ❗ the lowest.
-
-Kept rows per encoder x decoder (after the decoders' post-processing):
-
-| Encoder \ Decoder | Granite-4.2-30B-Nvfp4 [0] | Ornith-1.5-35B-A3B-Nvfp4 [1] | Llama-3.3-70B-Instruct-Nvfp4 [2] | Qwen3.8-27B-Awq-Int4 [3] | Mistral-Small-4-119B-2603-Nvfp4 [4] | Gemma-4-31B-It-Awq-4Bit [5] | Laguna-S-2.1-Nvfp4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 279 | 287 | 298 | 288 | 280 | 299 | 286 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 273 | 294 | 296 | 296 | 289 | 293 | 282 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 268 | 288 | 298 | 291 | 289 | 297 | 282 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 279 | 297 | 297 | 296 | 294 | 296 | 289 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 267 | 291 | 296 | 293 | 293 | 297 | 283 |
-| gemma-4-31B-it-AWQ-4bit [5] | 282 | 298 | 297 | 296 | 290 | 295 | 285 |
-| Laguna-S-2.1-NVFP4 [6] | 273 | 288 | 294 | 290 | 292 | 298 | 279 |
-| claude-sonnet-5 [8] | 278 | 290 | 299 | 292 | 293 | 296 | 290 |
-| gpt-5.6-terra [9] | 283 | 292 | 296 | 292 | 296 | 295 | 285 |
-
-Decoder post-processing:
-- Decoder 0 (`granite-4.2-30b-nvfp4`): 2,482 kept / 218 trashed of 2,700; failed requests 5; runtime 1.5 h (empty or too short: 6, refusal: 63, unfilled placeholder: 48, task meta-commentary: 137)
-- Decoder 1 (`Ornith-1.5-35B-A3B-NVFP4`): 2,625 kept / 75 trashed of 2,700; failed requests 3; runtime 0.6 h (empty or too short: 3, refusal: 10, unfilled placeholder: 24, task meta-commentary: 40)
-- Decoder 2 (`Llama-3.3-70B-Instruct-NVFP4`): 2,671 kept / 29 trashed of 2,700; failed requests 2; runtime 3.0 h (empty or too short: 2, refusal: 2, unfilled placeholder: 25)
-- Decoder 3 (`Qwen3.8-27B-AWQ-INT4`): 2,634 kept / 66 trashed of 2,700; failed requests 3; runtime 2.4 h (empty or too short: 5, unfilled placeholder: 56, task meta-commentary: 1, identical to source: 4)
-- Decoder 4 (`Mistral-Small-4-119B-2603-NVFP4`): 2,616 kept / 84 trashed of 2,700; failed requests 27; runtime 0.6 h (empty or too short: 28, refusal: 1, filler output: 1, unfilled placeholder: 51, task meta-commentary: 4)
-- Decoder 5 (`gemma-4-31B-it-AWQ-4bit`): 2,666 kept / 34 trashed of 2,700; failed requests 0; runtime 2.0 h (unfilled placeholder: 31, identical to source: 3)
-- Decoder 6 (`Laguna-S-2.1-NVFP4`): 2,561 kept / 139 trashed of 2,700; failed requests 0; runtime 1.0 h (empty or too short: 3, refusal: 41, filler output: 1, unfilled placeholder: 71, task meta-commentary: 50)
-
-## Rows filtered out per encoder
-
-Decoded rows that the decoders' post-processing rejected, per encoder (summed over decoders). A rejected row may carry several reasons, so the reason columns can add up to more than `Trashed`. `Empty Encodings At Encode` counts the encoder's own failed requests (those rows were never sent to a decoder).
-
-| Encoder | Decoded | Kept | Trashed | Trashed Rate | Empty Or Too Short | Filler Output | Identical To Source | Refusal | Task Meta-Commentary | Unfilled Placeholder | Empty Encodings At Encode |
-|---|---|---|---|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 2,100 | 2,017 | 83 | 0.0395 | 15 | 0 | 0 | 12 | 26 | 37 | 0 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 2,100 | 2,023 | 77 | 0.0367 | 7 | 0 | 2 | 17 | 32 | 31 | 0 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 2,100 | 2,013 | 87 | 0.0414 | 11 | 0 | 0 | 15 | 29 | 38 | 0 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 2,100 | 2,048 | 52 | 0.0248 | 0 | 0 | 1 | 15 | 22 | 24 | 0 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 2,100 | 2,020 | 80 | 0.0381 | 0 | 0 | 0 | 20 | 29 | 42 | 0 |
-| gemma-4-31B-it-AWQ-4bit [5] | 2,100 | 2,043 | 57 | 0.0271 | 1 | 0 | 0 | 7 | 19 | 36 | 0 |
-| Laguna-S-2.1-NVFP4 [6] | 2,100 | 2,014 | 86 | 0.0410 | 12 | 0 | 1 | 13 | 31 | 36 | 0 |
-| claude-sonnet-5 [8] | 2,100 | 2,038 | 62 | 0.0295 | 0 | 0 | 2 | 8 | 27 | 28 | 4 |
-| gpt-5.6-terra [9] | 2,100 | 2,039 | 61 | 0.0290 | 1 | 2 | 1 | 10 | 17 | 34 | 0 |
-
-Trashed rows per encoder x decoder:
-
-| Encoder \ Decoder | Granite-4.2-30B-Nvfp4 [0] | Ornith-1.5-35B-A3B-Nvfp4 [1] | Llama-3.3-70B-Instruct-Nvfp4 [2] | Qwen3.8-27B-Awq-Int4 [3] | Mistral-Small-4-119B-2603-Nvfp4 [4] | Gemma-4-31B-It-Awq-4Bit [5] | Laguna-S-2.1-Nvfp4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 21 | 13 | 2 | 12 | 20 | 1 | 14 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 27 | 6 | 4 | 4 | 11 | 7 | 18 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 32 | 12 | 2 | 9 | 11 | 3 | 18 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 21 | 3 | 3 | 4 | 6 | 4 | 11 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 33 | 9 | 4 | 7 | 7 | 3 | 17 |
-| gemma-4-31B-it-AWQ-4bit [5] | 18 | 2 | 3 | 4 | 10 | 5 | 15 |
-| Laguna-S-2.1-NVFP4 [6] | 27 | 12 | 6 | 10 | 8 | 2 | 21 |
-| claude-sonnet-5 [8] | 22 | 10 | 1 | 8 | 7 | 4 | 10 |
-| gpt-5.6-terra [9] | 17 | 8 | 4 | 8 | 4 | 5 | 15 |
-
-## Contents
-
-1. [EditLens score (higher = more AI-like) (`final_response_editlens_score_roberta_large`)](#editlens-score-higher-more-ai-like-final-response-editlens-score-roberta-large)
-2. [EditLens bucket (higher = more AI-like) (`final_response_editlens_bucket_roberta_large`)](#editlens-bucket-higher-more-ai-like-final-response-editlens-bucket-roberta-large)
-3. [Jaccard-1 distance to the source (`jaccard_1`)](#jaccard-1-distance-to-the-source-jaccard-1)
-4. [Jaccard-2 distance to the source (`jaccard_2`)](#jaccard-2-distance-to-the-source-jaccard-2)
-5. [Levenshtein distance to the source (`levenshtein`)](#levenshtein-distance-to-the-source-levenshtein)
-6. [Soft n-gram distance to the source (`softngram`)](#soft-n-gram-distance-to-the-source-softngram)
-7. [Embedding cosine distance to the source (`cosdist`)](#embedding-cosine-distance-to-the-source-cosdist)
-8. [BERTScore distance to the source (`bertscore`)](#bertscore-distance-to-the-source-bertscore)
-9. [BERTScore precision distance (`bertscore_precision`)](#bertscore-precision-distance-bertscore-precision)
-10. [BERTScore recall distance (`bertscore_recall`)](#bertscore-recall-distance-bertscore-recall)
-11. [MoverScore distance to the source (`moverscore`)](#moverscore-distance-to-the-source-moverscore)
-12. [Reranker distance to the source (`reranker`)](#reranker-distance-to-the-source-reranker)
-
-## Statistics
-
-### EditLens score (higher = more AI-like) (`final_response_editlens_score_roberta_large`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.6846 | 0.3230 | 0.6851 | 0.0519 | 2,017 | 7 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6301 | 0.3516 | 0.6311 | 0.0594 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8020 | 0.2808 | 0.8029 | 0.0537 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.7206 | 0.3045 | 0.7214 | 0.0604 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.7502 | 0.2992 | 0.7510 | 0.0507 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.7601 | 0.2859 | 0.7610 | 0.0541 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.7297 | 0.3202 | 0.7304 | 0.0581 | 2,014 | 7 |
-| ❗ claude-sonnet-5 [8] | 0.6165 | 0.3320 | 0.6172 | 0.0596 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 0.6188 | 0.3377 | 0.6196 | 0.0723 | 2,039 | 7 |
-
-![EditLens score (higher = more AI-like) per encoder](ENC_final_response_editlens_score_roberta_large.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.7294 | 0.5812 | 0.6424 | 0.7017 | 0.7140 | 0.6861 | 0.7409 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6730 | 0.5202 | 0.5839 | 0.6387 | 0.6840 | 0.6143 | 0.7037 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8380 | 0.7159 | 0.7351 | 0.8100 | 0.8507 | 0.8008 | 0.8698 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.7793 | 0.6096 | 0.7075 | 0.7118 | 0.7831 | 0.6767 | 0.7815 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.7917 | 0.6549 | 0.7240 | 0.7458 | 0.7837 | 0.7353 | 0.8218 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.8142 | 0.6625 | 0.7217 | 0.7729 | 0.8030 | 0.7314 | 0.8210 |
-| Laguna-S-2.1-NVFP4 [6] | 0.7863 | 0.6075 | 0.7137 | 0.7331 | 0.7743 | 0.7134 | 0.7847 |
-| claude-sonnet-5 [8] | 0.6738 | 0.5366 | 0.5582 | 0.6089 | 0.6620 | 0.5751 | 0.7058 |
-| gpt-5.6-terra [9] | 0.7041 | 0.5000 | 0.5527 | 0.6327 | 0.6510 | 0.5853 | 0.7111 |
-
-![EditLens score (higher = more AI-like) encoder x decoder](MATRIX_final_response_editlens_score_roberta_large.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.7540 | 0.2860 | 0.7544 | 0.0576 | 2,482 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5987 | 0.3359 | 0.5987 | 0.0677 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6598 | 0.3539 | 0.6599 | 0.0720 | 2,671 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.7061 | 0.3267 | 0.7062 | 0.0641 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.7450 | 0.3006 | 0.7451 | 0.0656 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.6800 | 0.3115 | 0.6798 | 0.0713 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.7709 | 0.2979 | 0.7712 | 0.0562 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5171 | 0.3285 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.8157 | 0.2599 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.6117 | 0.3315 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.8433 | 0.2657 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.6153 | 0.3220 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.8357 | 0.2405 | 3,168 |
-
-### EditLens bucket (higher = more AI-like) (`final_response_editlens_bucket_roberta_large`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 2.0347 | 1.1578 | 2.0366 | 0.1756 | 2,017 | 7 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 1.8586 | 1.2529 | 1.8620 | 0.1983 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 2.4302 | 1.0125 | 2.4331 | 0.1821 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 2.1699 | 1.1103 | 2.1729 | 0.2121 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 2.2540 | 1.0820 | 2.2569 | 0.1702 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 2.2746 | 1.0599 | 2.2775 | 0.1751 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 2.1787 | 1.1396 | 2.1809 | 0.1851 | 2,014 | 7 |
-| ❗ claude-sonnet-5 [8] | 1.7885 | 1.2023 | 1.7905 | 0.1856 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 1.8293 | 1.2224 | 1.8320 | 0.2319 | 2,039 | 7 |
-
-![EditLens bucket (higher = more AI-like) per encoder](ENC_final_response_editlens_bucket_roberta_large.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 2.2222 | 1.6899 | 1.8960 | 2.0903 | 2.1143 | 2.0268 | 2.2168 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 2.0073 | 1.4898 | 1.6993 | 1.9291 | 2.0000 | 1.7986 | 2.1099 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 2.5597 | 2.1146 | 2.2315 | 2.4570 | 2.5709 | 2.4310 | 2.6667 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 2.4050 | 1.7980 | 2.1313 | 2.1284 | 2.4014 | 1.9932 | 2.3529 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 2.3970 | 1.9381 | 2.1520 | 2.2389 | 2.3652 | 2.2121 | 2.4947 |
-| gemma-4-31B-it-AWQ-4bit [5] | 2.4468 | 1.9866 | 2.1347 | 2.3311 | 2.4483 | 2.1424 | 2.4526 |
-| Laguna-S-2.1-NVFP4 [6] | 2.3516 | 1.7951 | 2.1327 | 2.1793 | 2.3425 | 2.1141 | 2.3513 |
-| claude-sonnet-5 [8] | 1.9496 | 1.5414 | 1.6020 | 1.7740 | 1.9249 | 1.6622 | 2.0793 |
-| gpt-5.6-terra [9] | 2.1060 | 1.4555 | 1.6149 | 1.8425 | 1.9324 | 1.7356 | 2.1368 |
-
-![EditLens bucket (higher = more AI-like) encoder x decoder](MATRIX_final_response_editlens_bucket_roberta_large.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 2.2703 | 1.0460 | 2.2717 | 0.1992 | 2,482 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 1.7566 | 1.2066 | 1.7566 | 0.2185 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 1.9547 | 1.2913 | 1.9549 | 0.2402 | 2,671 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 2.1078 | 1.1585 | 2.1078 | 0.2131 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 2.2332 | 1.0787 | 2.2333 | 0.2290 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 2.0135 | 1.1342 | 2.0129 | 0.2328 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 2.3171 | 1.0708 | 2.3179 | 0.1875 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 1.4823 | 1.1974 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 2.4486 | 0.9586 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 1.7988 | 1.2034 | 3,320 |
-| Translate the entirety of this text into a sequence of emojis that captures the literal me... | 2.5443 | 0.9475 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 1.8138 | 1.1971 | 3,120 |
-| ✔️ Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 2.5540 | 0.8906 | 3,168 |
-
-### Jaccard-1 distance to the source (`jaccard_1`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.7666 | 0.1944 | 0.7668 | 0.0181 | 2,017 | 7 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7183 | 0.2432 | 0.7186 | 0.0217 | 2,023 | 7 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8339 | 0.1220 | 0.8341 | 0.0155 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.7944 | 0.1577 | 0.7946 | 0.0180 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8043 | 0.1521 | 0.8045 | 0.0170 | 2,020 | 7 |
-| ✔️ gemma-4-31B-it-AWQ-4bit [5] | 0.8342 | 0.0937 | 0.8345 | 0.0141 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.8019 | 0.1674 | 0.8020 | 0.0158 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.7543 | 0.1761 | 0.7545 | 0.0214 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 0.7440 | 0.1902 | 0.7442 | 0.0234 | 2,039 | 7 |
-
-![Jaccard-1 distance to the source per encoder](ENC_jaccard_1.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.7779 | 0.7590 | 0.7687 | 0.7339 | 0.7754 | 0.7564 | 0.7960 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7232 | 0.6976 | 0.7210 | 0.6857 | 0.7427 | 0.7090 | 0.7514 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8438 | 0.8245 | 0.8271 | 0.8088 | 0.8518 | 0.8276 | 0.8551 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.8066 | 0.7799 | 0.7962 | 0.7673 | 0.8183 | 0.7798 | 0.8141 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8099 | 0.7971 | 0.8019 | 0.7750 | 0.8233 | 0.7947 | 0.8294 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.8435 | 0.8280 | 0.8259 | 0.8173 | 0.8485 | 0.8208 | 0.8574 |
-| Laguna-S-2.1-NVFP4 [6] | 0.8107 | 0.7873 | 0.7983 | 0.7777 | 0.8215 | 0.7961 | 0.8227 |
-| claude-sonnet-5 [8] | 0.7634 | 0.7388 | 0.7506 | 0.7260 | 0.7791 | 0.7357 | 0.7876 |
-| gpt-5.6-terra [9] | 0.7644 | 0.7235 | 0.7449 | 0.7155 | 0.7578 | 0.7211 | 0.7824 |
-
-![Jaccard-1 distance to the source encoder x decoder](MATRIX_jaccard_1.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.7936 | 0.1682 | 0.7937 | 0.0376 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7707 | 0.1821 | 0.7706 | 0.0421 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.7816 | 0.1603 | 0.7816 | 0.0351 | 2,671 |
-| ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.7564 | 0.2014 | 0.7563 | 0.0415 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8021 | 0.1601 | 0.8020 | 0.0372 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.7713 | 0.1807 | 0.7712 | 0.0405 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.8106 | 0.1635 | 0.8107 | 0.0329 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.6064 | 0.2348 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.8594 | 0.0577 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.6784 | 0.1917 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9076 | 0.0519 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.7929 | 0.0865 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.8824 | 0.0314 | 3,168 |
-
-### Jaccard-2 distance to the source (`jaccard_2`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.8843 | 0.1891 | 0.8844 | 0.0147 | 2,017 | 7 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.8327 | 0.2423 | 0.8330 | 0.0165 | 2,023 | 7 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9398 | 0.1080 | 0.9400 | 0.0121 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.9153 | 0.1448 | 0.9155 | 0.0126 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9184 | 0.1380 | 0.9185 | 0.0139 | 2,020 | 7 |
-| ✔️ gemma-4-31B-it-AWQ-4bit [5] | 0.9503 | 0.0738 | 0.9504 | 0.0079 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.9114 | 0.1587 | 0.9115 | 0.0123 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.8826 | 0.1663 | 0.8827 | 0.0156 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 0.8743 | 0.1813 | 0.8745 | 0.0169 | 2,039 | 7 |
-
-![Jaccard-2 distance to the source per encoder](ENC_jaccard_2.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.8931 | 0.8801 | 0.8877 | 0.8569 | 0.8932 | 0.8741 | 0.9058 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.8333 | 0.8143 | 0.8388 | 0.8083 | 0.8527 | 0.8279 | 0.8553 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9484 | 0.9343 | 0.9365 | 0.9194 | 0.9549 | 0.9318 | 0.9543 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.9256 | 0.9073 | 0.9184 | 0.8964 | 0.9333 | 0.9024 | 0.9249 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9245 | 0.9149 | 0.9168 | 0.8931 | 0.9340 | 0.9095 | 0.9370 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.9560 | 0.9484 | 0.9457 | 0.9438 | 0.9572 | 0.9387 | 0.9629 |
-| Laguna-S-2.1-NVFP4 [6] | 0.9183 | 0.9005 | 0.9095 | 0.8918 | 0.9281 | 0.9065 | 0.9258 |
-| claude-sonnet-5 [8] | 0.8880 | 0.8712 | 0.8808 | 0.8627 | 0.9053 | 0.8679 | 0.9032 |
-| gpt-5.6-terra [9] | 0.8910 | 0.8592 | 0.8786 | 0.8530 | 0.8860 | 0.8559 | 0.8977 |
-
-![Jaccard-2 distance to the source encoder x decoder](MATRIX_jaccard_2.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.9086 | 0.1597 | 0.9087 | 0.0352 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.8923 | 0.1737 | 0.8922 | 0.0387 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9014 | 0.1487 | 0.9014 | 0.0313 | 2,671 |
-| ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.8806 | 0.1985 | 0.8806 | 0.0379 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9161 | 0.1464 | 0.9161 | 0.0325 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.8906 | 0.1712 | 0.8905 | 0.0345 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.9185 | 0.1528 | 0.9185 | 0.0308 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.7536 | 0.2503 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.9664 | 0.0315 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.8263 | 0.2107 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9851 | 0.0339 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.9147 | 0.0721 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.9794 | 0.0109 | 3,168 |
-
-### Levenshtein distance to the source (`levenshtein`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 2952.2920 | 2844.1391 | 2953.7796 | 168.8712 | 2,017 | 7 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 2827.0030 | 3803.2615 | 2830.1266 | 244.0772 | 2,023 | 7 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 3041.2966 | 2908.2688 | 3043.3284 | 108.4796 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 3019.2661 | 3429.4367 | 3020.2663 | 179.7201 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 3190.0277 | 4910.5124 | 3198.6141 | 303.7554 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 3051.9711 | 2984.0946 | 3053.2435 | 153.6386 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 3182.2925 | 4416.7275 | 3184.1194 | 313.8146 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 3151.3440 | 6250.0315 | 3155.4410 | 400.8880 | 2,038 | 7 |
-| ✔️ gpt-5.6-terra [9] | 3331.1275 | 7092.4672 | 3330.6634 | 572.0851 | 2,039 | 7 |
-
-![Levenshtein distance to the source per encoder](ENC_levenshtein.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 2953.7455 | 3216.3763 | 2986.4799 | 2707.5521 | 2996.9893 | 2730.3913 | 3084.9231 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 2828.2711 | 2997.6054 | 2721.8142 | 2521.1520 | 2872.8685 | 2575.3413 | 3293.8333 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 3117.5784 | 3013.7917 | 3045.7785 | 2909.7766 | 3146.7716 | 2881.0135 | 3188.5887 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 3053.2222 | 3190.7104 | 3023.1582 | 2759.2264 | 3140.6122 | 2751.6993 | 3223.2353 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 3650.9139 | 3079.7010 | 3107.6486 | 2875.1331 | 3265.9215 | 2812.0337 | 3598.9470 |
-| gemma-4-31B-it-AWQ-4bit [5] | 3021.1418 | 3069.5168 | 3059.4411 | 2884.8514 | 3282.7034 | 2825.5729 | 3229.4772 |
-| Laguna-S-2.1-NVFP4 [6] | 3095.0549 | 3077.5417 | 3601.1905 | 2760.0966 | 3329.3733 | 2829.6577 | 3595.9211 |
-| claude-sonnet-5 [8] | 3331.2626 | 3529.2345 | 2899.2040 | 2736.5993 | 3037.8259 | 2698.2399 | 3855.7207 |
-| gpt-5.6-terra [9] | 3209.4382 | 3291.9521 | 4520.0608 | 2753.8938 | 3160.7905 | 2710.4034 | 3668.1053 |
-
-![Levenshtein distance to the source encoder x decoder](MATRIX_levenshtein.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 3138.4533 | 4625.3624 | 3140.0698 | 225.7043 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 3162.6583 | 4945.6707 | 3162.9367 | 158.8055 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 3217.2280 | 6289.2638 | 3218.3084 | 510.8350 | 2,671 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 2767.4617 | 2841.2899 | 2767.5868 | 111.1271 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 3137.9687 | 2838.5017 | 3137.0952 | 139.3382 | 2,616 |
-| ❗ gemma-4-31B-it-AWQ-4bit [5] | 2757.4381 | 3004.1831 | 2757.1503 | 86.7623 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 3415.6892 | 5766.1180 | 3415.4168 | 251.8710 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 2330.5438 | 3133.3697 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 3106.8839 | 3308.2894 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 2682.8575 | 3286.1749 | 3,320 |
-| Translate the entirety of this text into a sequence of emojis that captures the literal me... | 3614.7831 | 7038.6579 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 3096.3061 | 5963.9306 | 3,120 |
-| ✔️ Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 3731.8021 | 2729.3455 | 3,168 |
-
-### Soft n-gram distance to the source (`softngram`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.7872 | 0.2739 | 0.7875 | 0.0263 | 2,017 | 7 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7288 | 0.3253 | 0.7293 | 0.0287 | 2,023 | 7 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8682 | 0.1900 | 0.8686 | 0.0242 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.8369 | 0.2298 | 0.8372 | 0.0237 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8350 | 0.2190 | 0.8353 | 0.0245 | 2,020 | 7 |
-| ✔️ gemma-4-31B-it-AWQ-4bit [5] | 0.8936 | 0.1406 | 0.8940 | 0.0210 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.8386 | 0.2332 | 0.8390 | 0.0234 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.7794 | 0.2492 | 0.7796 | 0.0292 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 0.7689 | 0.2679 | 0.7693 | 0.0308 | 2,039 | 7 |
-
-![Soft n-gram distance to the source per encoder](ENC_softngram.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.8069 | 0.7789 | 0.7853 | 0.7422 | 0.8016 | 0.7682 | 0.8297 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7447 | 0.7059 | 0.7229 | 0.6905 | 0.7621 | 0.7067 | 0.7723 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8829 | 0.8643 | 0.8531 | 0.8328 | 0.8976 | 0.8479 | 0.9016 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.8590 | 0.8199 | 0.8330 | 0.8044 | 0.8674 | 0.8143 | 0.8622 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8461 | 0.8192 | 0.8288 | 0.7986 | 0.8635 | 0.8187 | 0.8722 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.9090 | 0.8925 | 0.8724 | 0.8756 | 0.9130 | 0.8687 | 0.9265 |
-| Laguna-S-2.1-NVFP4 [6] | 0.8596 | 0.8181 | 0.8272 | 0.8085 | 0.8670 | 0.8234 | 0.8689 |
-| claude-sonnet-5 [8] | 0.7922 | 0.7655 | 0.7602 | 0.7501 | 0.8151 | 0.7485 | 0.8259 |
-| gpt-5.6-terra [9] | 0.8033 | 0.7441 | 0.7568 | 0.7360 | 0.7873 | 0.7394 | 0.8180 |
-
-![Soft n-gram distance to the source encoder x decoder](MATRIX_softngram.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.8337 | 0.2350 | 0.8337 | 0.0481 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.8010 | 0.2626 | 0.8009 | 0.0552 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8044 | 0.2268 | 0.8044 | 0.0473 | 2,671 |
-| ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.7821 | 0.2828 | 0.7821 | 0.0537 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8417 | 0.2289 | 0.8416 | 0.0490 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.7930 | 0.2504 | 0.7929 | 0.0513 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.8530 | 0.2245 | 0.8530 | 0.0441 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5681 | 0.3124 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.9337 | 0.0855 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.6387 | 0.2586 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9785 | 0.0750 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.8389 | 0.1364 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.9718 | 0.0417 | 3,168 |
-
-### Embedding cosine distance to the source (`cosdist`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.2722 | 0.2088 | 0.2720 | 0.0153 | 2,017 | 7 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.2825 | 0.2274 | 0.2825 | 0.0167 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.3846 | 0.2449 | 0.3843 | 0.0169 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.2912 | 0.1927 | 0.2912 | 0.0199 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.3194 | 0.2215 | 0.3192 | 0.0175 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.3311 | 0.1832 | 0.3312 | 0.0190 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.3353 | 0.2200 | 0.3352 | 0.0171 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.2420 | 0.1747 | 0.2419 | 0.0198 | 2,038 | 7 |
-| ❗ gpt-5.6-terra [9] | 0.2224 | 0.1900 | 0.2224 | 0.0183 | 2,039 | 7 |
-
-![Embedding cosine distance to the source per encoder](ENC_cosdist.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.2676 | 0.2554 | 0.3055 | 0.2731 | 0.2574 | 0.2712 | 0.2735 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.2767 | 0.2658 | 0.3139 | 0.2689 | 0.2951 | 0.2665 | 0.2905 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.3617 | 0.3618 | 0.4113 | 0.3872 | 0.3869 | 0.3816 | 0.3993 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.2888 | 0.2628 | 0.3298 | 0.2918 | 0.2983 | 0.2713 | 0.2955 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.2997 | 0.2950 | 0.3498 | 0.3205 | 0.3321 | 0.3127 | 0.3242 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.3337 | 0.3055 | 0.3686 | 0.3277 | 0.3351 | 0.3111 | 0.3363 |
-| Laguna-S-2.1-NVFP4 [6] | 0.3187 | 0.3087 | 0.3634 | 0.3382 | 0.3458 | 0.3265 | 0.3448 |
-| claude-sonnet-5 [8] | 0.2435 | 0.2190 | 0.2736 | 0.2280 | 0.2569 | 0.2168 | 0.2553 |
-| gpt-5.6-terra [9] | 0.2235 | 0.2043 | 0.2517 | 0.2086 | 0.2330 | 0.1975 | 0.2381 |
-
-![Embedding cosine distance to the source encoder x decoder](MATRIX_cosdist.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.2900 | 0.2050 | 0.2904 | 0.0411 | 2,482 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.2753 | 0.2129 | 0.2754 | 0.0456 | 2,625 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.3297 | 0.2252 | 0.3297 | 0.0469 | 2,671 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.2938 | 0.2093 | 0.2938 | 0.0529 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.3045 | 0.2164 | 0.3045 | 0.0470 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.2840 | 0.2050 | 0.2839 | 0.0533 | 2,666 |
-| Laguna-S-2.1-NVFP4 [6] | 0.3061 | 0.2143 | 0.3064 | 0.0472 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1231 | 0.0941 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.2712 | 0.1445 | 2,704 |
-| ❗ Reformat this text into a sensible, structured JSON object. | 0.1166 | 0.0757 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.6031 | 0.1830 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.3009 | 0.1374 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.4050 | 0.1465 | 3,168 |
-
-### BERTScore distance to the source (`bertscore`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1751 | 0.0675 | 0.1751 | 0.0062 | 2,017 | 7 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1639 | 0.0751 | 0.1639 | 0.0066 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.2012 | 0.0562 | 0.2013 | 0.0061 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1822 | 0.0565 | 0.1822 | 0.0067 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1871 | 0.0580 | 0.1872 | 0.0055 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1954 | 0.0435 | 0.1955 | 0.0053 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1896 | 0.0614 | 0.1896 | 0.0052 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.1678 | 0.0603 | 0.1678 | 0.0082 | 2,038 | 7 |
-| ❗ gpt-5.6-terra [9] | 0.1636 | 0.0641 | 0.1637 | 0.0094 | 2,039 | 7 |
-
-![BERTScore distance to the source per encoder](ENC_bertscore.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1790 | 0.1702 | 0.1801 | 0.1660 | 0.1745 | 0.1705 | 0.1852 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1630 | 0.1568 | 0.1679 | 0.1561 | 0.1706 | 0.1588 | 0.1743 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.2009 | 0.1956 | 0.2053 | 0.1925 | 0.2052 | 0.1979 | 0.2116 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1831 | 0.1752 | 0.1857 | 0.1745 | 0.1890 | 0.1758 | 0.1924 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1867 | 0.1816 | 0.1911 | 0.1797 | 0.1925 | 0.1832 | 0.1952 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1956 | 0.1905 | 0.1979 | 0.1907 | 0.1995 | 0.1893 | 0.2050 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1899 | 0.1825 | 0.1932 | 0.1840 | 0.1934 | 0.1863 | 0.1980 |
-| claude-sonnet-5 [8] | 0.1685 | 0.1605 | 0.1700 | 0.1601 | 0.1749 | 0.1584 | 0.1823 |
-| gpt-5.6-terra [9] | 0.1661 | 0.1561 | 0.1691 | 0.1537 | 0.1679 | 0.1525 | 0.1804 |
-
-![BERTScore distance to the source encoder x decoder](MATRIX_bertscore.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1813 | 0.0632 | 0.1814 | 0.0126 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1743 | 0.0604 | 0.1743 | 0.0137 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1845 | 0.0609 | 0.1845 | 0.0128 | 2,671 |
-| ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.1730 | 0.0638 | 0.1730 | 0.0139 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1853 | 0.0565 | 0.1853 | 0.0128 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1748 | 0.0592 | 0.1747 | 0.0149 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.1916 | 0.0686 | 0.1916 | 0.0115 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1192 | 0.0565 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.1962 | 0.0325 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.1354 | 0.0466 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.2512 | 0.0543 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.1774 | 0.0329 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.2150 | 0.0240 | 3,168 |
-
-### BERTScore precision distance (`bertscore_precision`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1760 | 0.0692 | 0.1761 | 0.0091 | 2,017 | 7 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1638 | 0.0763 | 0.1639 | 0.0099 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1982 | 0.0575 | 0.1983 | 0.0093 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1826 | 0.0579 | 0.1827 | 0.0102 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1871 | 0.0588 | 0.1872 | 0.0084 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1943 | 0.0467 | 0.1944 | 0.0089 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1892 | 0.0637 | 0.1893 | 0.0084 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.1697 | 0.0643 | 0.1697 | 0.0114 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 0.1678 | 0.0666 | 0.1679 | 0.0119 | 2,039 | 7 |
-
-![BERTScore precision distance per encoder](ENC_bertscore_precision.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1812 | 0.1685 | 0.1803 | 0.1647 | 0.1796 | 0.1665 | 0.1917 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1649 | 0.1538 | 0.1667 | 0.1541 | 0.1729 | 0.1538 | 0.1811 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1999 | 0.1898 | 0.2023 | 0.1855 | 0.2069 | 0.1905 | 0.2131 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1845 | 0.1726 | 0.1862 | 0.1720 | 0.1931 | 0.1717 | 0.1990 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1888 | 0.1784 | 0.1905 | 0.1774 | 0.1960 | 0.1791 | 0.2001 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1942 | 0.1862 | 0.1952 | 0.1889 | 0.2026 | 0.1829 | 0.2105 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1914 | 0.1797 | 0.1923 | 0.1808 | 0.1962 | 0.1810 | 0.2036 |
-| claude-sonnet-5 [8] | 0.1707 | 0.1605 | 0.1686 | 0.1626 | 0.1781 | 0.1555 | 0.1920 |
-| gpt-5.6-terra [9] | 0.1727 | 0.1574 | 0.1717 | 0.1579 | 0.1725 | 0.1528 | 0.1903 |
-
-![BERTScore precision distance encoder x decoder](MATRIX_bertscore_precision.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1831 | 0.0641 | 0.1831 | 0.0111 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1719 | 0.0607 | 0.1719 | 0.0121 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1837 | 0.0623 | 0.1838 | 0.0119 | 2,671 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1716 | 0.0661 | 0.1716 | 0.0117 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1886 | 0.0597 | 0.1886 | 0.0123 | 2,616 |
-| ❗ gemma-4-31B-it-AWQ-4bit [5] | 0.1705 | 0.0574 | 0.1704 | 0.0132 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.1979 | 0.0712 | 0.1979 | 0.0097 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1227 | 0.0605 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.2053 | 0.0380 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.1427 | 0.0534 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.2452 | 0.0601 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.1659 | 0.0352 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.2147 | 0.0288 | 3,168 |
-
-### BERTScore recall distance (`bertscore_recall`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1731 | 0.0709 | 0.1731 | 0.0043 | 2,017 | 7 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1630 | 0.0783 | 0.1630 | 0.0040 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.2033 | 0.0612 | 0.2033 | 0.0036 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1807 | 0.0613 | 0.1808 | 0.0033 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1862 | 0.0633 | 0.1862 | 0.0032 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1957 | 0.0479 | 0.1957 | 0.0025 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1888 | 0.0656 | 0.1888 | 0.0028 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.1650 | 0.0617 | 0.1650 | 0.0057 | 2,038 | 7 |
-| ❗ gpt-5.6-terra [9] | 0.1585 | 0.0668 | 0.1585 | 0.0071 | 2,039 | 7 |
-
-![BERTScore recall distance per encoder](ENC_bertscore_recall.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1756 | 0.1709 | 0.1787 | 0.1665 | 0.1685 | 0.1737 | 0.1777 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1601 | 0.1590 | 0.1682 | 0.1573 | 0.1673 | 0.1629 | 0.1664 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.2009 | 0.2004 | 0.2072 | 0.1985 | 0.2025 | 0.2043 | 0.2091 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.1808 | 0.1768 | 0.1840 | 0.1760 | 0.1842 | 0.1790 | 0.1846 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1837 | 0.1838 | 0.1906 | 0.1811 | 0.1882 | 0.1864 | 0.1893 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1961 | 0.1939 | 0.1997 | 0.1916 | 0.1954 | 0.1948 | 0.1984 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1875 | 0.1842 | 0.1930 | 0.1860 | 0.1897 | 0.1905 | 0.1906 |
-| claude-sonnet-5 [8] | 0.1653 | 0.1596 | 0.1704 | 0.1565 | 0.1710 | 0.1606 | 0.1713 |
-| gpt-5.6-terra [9] | 0.1586 | 0.1538 | 0.1655 | 0.1486 | 0.1625 | 0.1515 | 0.1692 |
-
-![BERTScore recall distance encoder x decoder](MATRIX_bertscore_recall.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.1786 | 0.0676 | 0.1787 | 0.0143 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1758 | 0.0656 | 0.1758 | 0.0154 | 2,625 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1842 | 0.0661 | 0.1842 | 0.0138 | 2,671 |
-| ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.1736 | 0.0668 | 0.1736 | 0.0163 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1810 | 0.0589 | 0.1810 | 0.0133 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.1782 | 0.0661 | 0.1782 | 0.0165 | 2,666 |
-| Laguna-S-2.1-NVFP4 [6] | 0.1840 | 0.0711 | 0.1841 | 0.0135 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1149 | 0.0574 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.1861 | 0.0351 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.1272 | 0.0459 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.2560 | 0.0542 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.1876 | 0.0424 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.2146 | 0.0290 | 3,168 |
-
-### MoverScore distance to the source (`moverscore`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.6318 | 0.1461 | 0.6318 | 0.0138 | 2,017 | 7 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6028 | 0.1718 | 0.6030 | 0.0155 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6854 | 0.1031 | 0.6854 | 0.0120 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.6506 | 0.1194 | 0.6507 | 0.0134 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.6590 | 0.1166 | 0.6591 | 0.0122 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.6786 | 0.0804 | 0.6787 | 0.0103 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | 0.6628 | 0.1256 | 0.6629 | 0.0106 | 2,014 | 7 |
-| claude-sonnet-5 [8] | 0.6204 | 0.1298 | 0.6204 | 0.0169 | 2,038 | 7 |
-| gpt-5.6-terra [9] | 0.6116 | 0.1394 | 0.6117 | 0.0201 | 2,039 | 7 |
-
-![MoverScore distance to the source per encoder](ENC_moverscore.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.6398 | 0.6236 | 0.6445 | 0.6099 | 0.6320 | 0.6206 | 0.6524 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6014 | 0.5857 | 0.6152 | 0.5827 | 0.6191 | 0.5925 | 0.6242 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6869 | 0.6755 | 0.6945 | 0.6657 | 0.6945 | 0.6783 | 0.7026 |
-| Qwen3.8-27B-AWQ-INT4 [3] | 0.6543 | 0.6371 | 0.6604 | 0.6338 | 0.6646 | 0.6370 | 0.6678 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.6596 | 0.6495 | 0.6692 | 0.6399 | 0.6701 | 0.6497 | 0.6757 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.6788 | 0.6698 | 0.6838 | 0.6698 | 0.6859 | 0.6658 | 0.6970 |
-| Laguna-S-2.1-NVFP4 [6] | 0.6624 | 0.6487 | 0.6714 | 0.6501 | 0.6721 | 0.6572 | 0.6781 |
-| claude-sonnet-5 [8] | 0.6229 | 0.6052 | 0.6277 | 0.6033 | 0.6350 | 0.6003 | 0.6487 |
-| gpt-5.6-terra [9] | 0.6207 | 0.5944 | 0.6271 | 0.5884 | 0.6202 | 0.5878 | 0.6436 |
-
-![MoverScore distance to the source encoder x decoder](MATRIX_moverscore.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | 0.6473 | 0.1266 | 0.6474 | 0.0268 | 2,482 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6322 | 0.1336 | 0.6322 | 0.0303 | 2,625 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6549 | 0.1239 | 0.6549 | 0.0260 | 2,671 |
-| ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.6271 | 0.1450 | 0.6271 | 0.0305 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | 0.6549 | 0.1188 | 0.6548 | 0.0270 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | 0.6322 | 0.1299 | 0.6321 | 0.0315 | 2,666 |
-| ✔️ Laguna-S-2.1-NVFP4 [6] | 0.6655 | 0.1334 | 0.6656 | 0.0242 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| ❗ Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5136 | 0.1551 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.6846 | 0.0623 | 2,704 |
-| Reformat this text into a sensible, structured JSON object. | 0.5584 | 0.1240 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.7634 | 0.0837 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.6543 | 0.0692 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.7142 | 0.0390 | 3,168 |
-
-### Reranker distance to the source (`reranker`)
-
-Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation.
-
-| Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders |
-|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | -1.9931 | 6.0803 | -2.0271 | 2.1146 | 2,017 | 7 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | -1.7876 | 6.4575 | -1.8051 | 1.8821 | 2,023 | 7 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 1.8343 | 6.7894 | 1.7986 | 1.9136 | 2,013 | 7 |
-| Qwen3.8-27B-AWQ-INT4 [3] | -1.6688 | 5.8839 | -1.6815 | 2.1851 | 2,048 | 7 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | -0.4640 | 6.3430 | -0.4909 | 2.0717 | 2,020 | 7 |
-| gemma-4-31B-it-AWQ-4bit [5] | -0.6061 | 5.8655 | -0.6185 | 2.2919 | 2,043 | 7 |
-| Laguna-S-2.1-NVFP4 [6] | -0.4634 | 6.3073 | -0.4826 | 1.9459 | 2,014 | 7 |
-| claude-sonnet-5 [8] | -2.7594 | 5.5749 | -2.7809 | 2.0750 | 2,038 | 7 |
-| ❗ gpt-5.6-terra [9] | -3.4418 | 5.7955 | -3.4548 | 2.1491 | 2,039 | 7 |
-
-![Reranker distance to the source per encoder](ENC_reranker.png)
-
-Encoder x decoder cell means:
-
-| Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] |
-|---|---|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | -2.9872 | -3.8632 | 2.9989 | -2.3861 | -3.1653 | -2.2301 | -2.5569 |
-| Ornith-1.5-35B-A3B-NVFP4 [1] | -2.8964 | -3.4596 | 2.6671 | -2.3179 | -2.2910 | -2.3973 | -1.9409 |
-| Llama-3.3-70B-Instruct-NVFP4 [2] | 0.4150 | 0.0829 | 6.2713 | 1.3993 | 1.0511 | 1.8009 | 1.5697 |
-| Qwen3.8-27B-AWQ-INT4 [3] | -2.8290 | -3.7474 | 3.4932 | -2.0536 | -2.2046 | -2.4160 | -2.0132 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | -1.7513 | -2.5116 | 4.3805 | -0.9309 | -0.8193 | -0.8764 | -0.9273 |
-| gemma-4-31B-it-AWQ-4bit [5] | -1.4618 | -2.8054 | 4.8353 | -1.0130 | -1.2538 | -1.2923 | -1.3384 |
-| Laguna-S-2.1-NVFP4 [6] | -1.6698 | -2.3852 | 4.0789 | -0.7677 | -1.0757 | -0.9518 | -0.6068 |
-| claude-sonnet-5 [8] | -3.3729 | -4.7251 | 2.1394 | -3.7727 | -3.0359 | -3.5779 | -3.1211 |
-| gpt-5.6-terra [9] | -4.1568 | -5.1910 | 1.7251 | -4.3837 | -3.9796 | -4.2069 | -3.9907 |
-
-![Reranker distance to the source encoder x decoder](MATRIX_reranker.png)
-
-Per decoder, marginalised over every encoder (for contrast):
-
-| Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N |
-|---|---|---|---|---|---|
-| granite-4.2-30b-nvfp4 [0] | -2.3174 | 5.6882 | -2.3011 | 1.2684 | 2,482 |
-| ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | -3.1828 | 5.7025 | -3.1784 | 1.4592 | 2,625 |
-| ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 3.6210 | 5.9929 | 3.6211 | 1.3494 | 2,671 |
-| Qwen3.8-27B-AWQ-INT4 [3] | -1.8037 | 5.9939 | -1.8029 | 1.6302 | 2,634 |
-| Mistral-Small-4-119B-2603-NVFP4 [4] | -1.8647 | 6.2847 | -1.8638 | 1.4308 | 2,616 |
-| gemma-4-31B-it-AWQ-4bit [5] | -1.7910 | 5.8269 | -1.7942 | 1.6557 | 2,666 |
-| Laguna-S-2.1-NVFP4 [6] | -1.6686 | 6.1229 | -1.6584 | 1.5186 | 2,561 |
-
-Per encoding instruction, marginalised over every encoder and decoder:
-
-| Encoding instruction | Mean | Std | N |
-|---|---|---|---|
-| Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | -5.2104 | 3.6907 | 3,108 |
-| Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | -2.3215 | 4.9092 | 2,704 |
-| ❗ Reformat this text into a sensible, structured JSON object. | -5.5315 | 3.4234 | 3,320 |
-| ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 7.0536 | 6.1041 | 2,835 |
-| Write a detailed descriptor for how this text is stylistically differentiated from other t... | -2.0842 | 4.5451 | 3,120 |
-| Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 1.3351 | 4.9330 | 3,168 |
-
+# WIP encoder/decoder trial
+Waiting for statistics to finish generating...

runs/shard_4.json

@@ -6,69 +6,72 @@
   "reasoning_effort": "none"
  },
  "ignored_params": {},
- "prompt_tokens": 2635892,
- "completion_tokens": 1862415,
- "failed_requests": 27,
- "runtime_seconds": 2172.9,
+ "prompt_tokens": 4593639,
+ "completion_tokens": 3870838,
+ "failed_requests": 20,
+ "runtime_seconds": 3954.3,
  "stage": "decode",
  "decoder_index": 4,
- "rows_in": 2700,
- "rows_kept": 2616,
- "rows_trashed": 84,
+ "rows_in": 6296,
+ "rows_kept": 6203,
+ "rows_trashed": 93,
  "dropped": {
-  "empty or too short": 28,
-  "refusal": 1,
-  "filler output": 1,
-  "unfilled placeholder": 51,
-  "task meta-commentary": 4,
+  "empty or too short": 21,
+  "refusal": 0,
+  "filler output": 15,
+  "unfilled placeholder": 42,
+  "task meta-commentary": 5,
   "echoed instruction": 0,
-  "identical to source": 0
+  "identical to source": 10
  },
  "per_encoder": {
   "0": {
-   "kept": 280,
-   "trashed": 20,
+   "kept": 967,
+   "trashed": 33,
    "skipped_empty_encoding": 0
   },
   "1": {
-   "kept": 289,
-   "trashed": 11,
+   "kept": 973,
+   "trashed": 27,
    "skipped_empty_encoding": 0
   },
   "2": {
-   "kept": 289,
-   "trashed": 11,
+   "kept": 979,
+   "trashed": 21,
    "skipped_empty_encoding": 0
   },
   "3": {
-   "kept": 294,
-   "trashed": 6,
+   "kept": 984,
+   "trashed": 16,
    "skipped_empty_encoding": 0
   },
   "4": {
-   "kept": 293,
-   "trashed": 7,
+   "kept": 988,
+   "trashed": 12,
    "skipped_empty_encoding": 0
   },
   "5": {
-   "kept": 290,
-   "trashed": 10,
+   "kept": 983,
+   "trashed": 17,
    "skipped_empty_encoding": 0
   },
   "6": {
-   "kept": 292,
-   "trashed": 8,
+   "kept": 978,
+   "trashed": 22,
    "skipped_empty_encoding": 0
   },
   "8": {
-   "kept": 293,
-   "trashed": 7,
-   "skipped_empty_encoding": 0
+   "kept": 979,
+   "trashed": 17,
+   "skipped_empty_encoding": 4
   },
   "9": {
-   "kept": 296,
-   "trashed": 4,
+   "kept": 988,
+   "trashed": 12,
    "skipped_empty_encoding": 0
   }
- }
+ },
+ "rows_per_encoder": 1000,
+ "total_kept": 8819,
+ "total_trashed": 177
 }