Kambo-v1 SQL + Code, DynQuant 4-bit (4.25 bits per weight, packed)

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1,15 @@
+Kambo-v1
+Copyright 2026 Vikrampal Kamboj
+
+This product is licensed under the Apache License, Version 2.0 (see LICENSE).
+
+The following files are third-party material, also distributed under the
+Apache License, Version 2.0:
+
+    tokenizer.json
+    tokenizer_config.json
+
+MODIFICATIONS. These files have been modified from their original form. The
+vocabulary and merge tables are unchanged; the accompanying configuration was
+modified to set the chat template, the end-of-turn and padding tokens, and the
+maximum sequence length used by this model.

README.md

@@ -0,0 +1,288 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model: VikramPal/kambo-v1-sql-code
+base_model_relation: quantized
+datasets:
+- gretelai/synthetic_text_to_sql
+- Salesforce/wikisql
+- b-mc2/sql-create-context
+- nvidia/OpenCodeInstruct
+tags:
+- dynquant
+- quantized
+- 4-bit
+- text-to-sql
+- code
+- mixture-of-experts
+- moe
+- hybrid-architecture
+---
+
+# Kambo-v1 SQL + Code, DynQuant 4-bit
+
+[VikramPal/kambo-v1-sql-code](https://huggingface.co/VikramPal/kambo-v1-sql-code) quantized with [DynQuant](https://github.com/kambojvikram/dynquant) to 4.25 bits per parameter, scales included: the same byte budget as uniform 4-bit, spent unevenly. 0.836 GiB of weights instead of 3.150 GiB. The other width: [3-bit](https://huggingface.co/VikramPal/kambo-v1-sql-code-DynQuant-3bit).
+
+Against the bf16 fine-tune this checkpoint loses 4.85 points on text-to-SQL (49.06% against 53.91%, separated after Holm correction; by source (exploratory rows) Gretel -3.06, WikiSQL -10.76 and Spider dev -0.73, so WikiSQL accounts for 74% of the net loss in items) and is not separated from the fine-tune on HumanEval (+0.61) and MBPP (-0.60). Beside the unquantized base model it scores higher on text-to-SQL (49.06% against 42.87%) and MBPP (28.20% against 25.40%), and the same on HumanEval (31.10%); these are descriptions, not planned tests. At 0.13% fewer bytes than uniform 4-bit, it beats that control on text-to-SQL (+5.30), and is not separated from it on HumanEval (+6.71) and MBPP (+4.40, uncorrected p = 0.0115). On held-out loss (KL to the fine-tune, paired by conversation) it is closer to the fine-tune than uniform 4-bit (|z| = 23.0) and one draw of the permuted-signal null (|z| = 13.0).
+
+## What this is
+
+| | |
+|---|---|
+| quantized from | [VikramPal/kambo-v1-sql-code](https://huggingface.co/VikramPal/kambo-v1-sql-code), the bf16 fine-tune of [VikramPal/kambo-v1](https://huggingface.co/VikramPal/kambo-v1) |
+| method | DynQuant 0.5.3, per-matrix widths of 2, 3, 4 or 8 bits from the fine-tune's own training signal, asymmetric, groups of 128 |
+| size | 0.836 GiB of weights (4.2479 bits per parameter, scales and the bf16 remainder included); 898,079,352 bytes of safetensors |
+| memory | 0.836 GiB resident on the GPU after loading (weights and buffers, before any KV cache), 100.0% of the map's prediction; 0.836 GiB on disk |
+| loads with | `transformers` + `dynquant`, `trust_remote_code=True`, **bfloat16 only**. The usage snippet below ran against this repo's files with dynquant 0.5.3 under transformers 5.14.1 (torch 2.13.0+cu130) and 5.18.0 (torch 2.14.1+cu130) |
+
+## Results
+
+| arm | text-to-SQL (2,454) | HumanEval (164) | MBPP (500) | weights | bits/param |
+|---|---:|---:|---:|---:|---:|
+| Kambo-v1 (base) | 42.87% (1052/2454) | 31.10% (51/164) | 25.40% (127/500) | 3.150 GiB | 16.0000 |
+| fine-tune, bf16 | 53.91% (1323/2454) | 30.49% (50/164) | 28.80% (144/500) | 3.150 GiB | 16.0000 |
+| **DynQuant 4-bit** (this repo) | 49.06% (1204/2454) | 31.10% (51/164) | 28.20% (141/500) | 0.836 GiB | 4.2479 |
+| uniform 4-bit | 43.77% (1074/2454) | 24.39% (40/164) | 23.80% (119/500) | 0.837 GiB | 4.2535 |
+| DynQuant 3-bit | 38.75% (951/2454) | 19.51% (32/164) | 20.00% (100/500) | 0.640 GiB | 3.2495 |
+| uniform 3-bit | 24.33% (597/2454) | 2.44% (4/164) | 6.40% (32/500) | 0.641 GiB | 3.2538 |
+
+Accuracy with the correct count. This arm was scored from this repo's packed checkpoint. The uniform arms put every quantized matrix at one width with the same quantizer and byte accounting. Each DynQuant arm's bytes are within 0.13% of its uniform control's.
+
+Text-to-SQL by source:
+
+| arm | Gretel test (a training source) | WikiSQL test (a training source) | Spider dev (not a training source; see below) |
+|---|---:|---:|---:|
+| Kambo-v1 (base) | 52.93% (433/818) | 51.71% (423/818) | 23.96% (196/818) |
+| fine-tune, bf16 | 60.15% (492/818) | 75.92% (621/818) | 25.67% (210/818) |
+| **DynQuant 4-bit** | 57.09% (467/818) | 65.16% (533/818) | 24.94% (204/818) |
+| uniform 4-bit | 50.73% (415/818) | 61.37% (502/818) | 19.19% (157/818) |
+| DynQuant 3-bit | 45.48% (372/818) | 56.72% (464/818) | 14.06% (115/818) |
+| uniform 3-bit | 28.24% (231/818) | 41.20% (337/818) | 3.55% (29/818) |
+
+Spider is not one of the three training sources, but sql-create-context, which is, was built partly from Spider. Training rows asking a Spider dev question (after folding case, punctuation and whitespace) were removed; other Spider-derived rows (Spider train questions, for instance) can be in the training mix.
+
+## How this arm compares
+
+McNemar exact over the per-item hits: every row pairs two arms on the same problems in the same order, so only the items the two arms disagree on (`+` won by the first arm, `−` by the second) carry information. Delta is the first arm minus the second, in points. The 95% interval is exact and conditional on the number of disagreements (Clopper–Pearson on the first arm's share of them, scaled by their share of the items), so it excludes zero exactly when the unadjusted p is below 0.05. `p (Holm)` is step-down corrected across the 18 planned tests that could be computed (18 were declared before any fine-tuned arm was scored: 6 comparisons × 3 tasks). `separated` means Holm p < 0.05; `not separated` means this test cannot tell the two arms apart, not that they are equal: the interval shows how large a difference remains possible.
+
+| comparison | task | first | second | delta (pts) | 95% CI | disagreements | p | p (Holm) | verdict |
+|---|---|---:|---:|---:|---|---:|---:|---:|---|
+| DynQuant 4-bit vs the bf16 fine-tune | text-to-SQL | 1204/2454 | 1323/2454 | -4.85 | [-6.22, -3.39] | +110 / −229 | 9.48e-11 | 1.33e-09 | separated |
+| DynQuant 4-bit vs the bf16 fine-tune | HumanEval | 51/164 | 50/164 | +0.61 | [-5.70, +6.77] | +13 / −12 | 1.00 | 1.00 | not separated |
+| DynQuant 4-bit vs the bf16 fine-tune | MBPP | 141/500 | 144/500 | -0.60 | [-3.30, +2.18] | +21 / −24 | 0.766 | 1.00 | not separated |
+| DynQuant 4-bit vs uniform 4-bit | text-to-SQL | 1204/2454 | 1074/2454 | +5.30 | [+3.68, +6.84] | +273 / −143 | 1.82e-10 | 2.36e-09 | separated |
+| DynQuant 4-bit vs uniform 4-bit | HumanEval | 51/164 | 40/164 | +6.71 | [-0.29, +12.28] | +20 / −9 | 0.0614 | 0.249 | not separated |
+| DynQuant 4-bit vs uniform 4-bit | MBPP | 141/500 | 119/500 | +4.40 | [+0.95, +7.46] | +46 / −24 | 0.0115 | 0.0744 | not separated |
+
+Secondary and exploratory rows, not corrected for multiplicity (the per-source rows are cuts of the text-to-SQL row with the same label, and the pooled-code row is the union of the two code rows with that label; neither is further evidence):
+
+| comparison | task | first | second | delta (pts) | 95% CI | disagreements | p |
+|---|---|---:|---:|---:|---|---:|---:|
+| DynQuant 4-bit vs the bf16 fine-tune | text-to-SQL / gretel | 467/818 | 492/818 | -3.06 | [-5.05, -0.83] | +27 / −52 | 0.00655 |
+| DynQuant 4-bit vs the bf16 fine-tune | text-to-SQL / wikisql | 533/818 | 621/818 | -10.76 | [-13.06, -8.03] | +32 / −120 | 3.53e-13 |
+| DynQuant 4-bit vs the bf16 fine-tune | text-to-SQL / spider | 204/818 | 210/818 | -0.73 | [-3.29, +1.86] | +51 / −57 | 0.631 |
+| DynQuant 4-bit vs the bf16 fine-tune | code (humaneval+mbpp) | 192/664 | 194/664 | -0.30 | [-2.86, +2.28] | +34 / −36 | 0.905 |
+| DynQuant 4-bit vs uniform 4-bit | text-to-SQL / gretel | 467/818 | 415/818 | +6.36 | [+3.44, +9.00] | +98 / −46 | 1.75e-05 |
+| DynQuant 4-bit vs uniform 4-bit | text-to-SQL / wikisql | 533/818 | 502/818 | +3.79 | [+0.75, +6.67] | +90 / −59 | 0.0137 |
+| DynQuant 4-bit vs uniform 4-bit | text-to-SQL / spider | 204/818 | 157/818 | +5.75 | [+3.05, +8.16] | +85 / −38 | 2.72e-05 |
+| DynQuant 4-bit vs uniform 4-bit | code (humaneval+mbpp) | 192/664 | 159/664 | +4.97 | [+1.93, +7.70] | +66 / −33 | 0.00119 |
+| 4-bit vs 3-bit | text-to-SQL | 1204/2454 | 951/2454 | +10.31 | [+8.77, +11.71] | +357 / −104 | 1.64e-33 |
+| 4-bit vs 3-bit | text-to-SQL / gretel | 467/818 | 372/818 | +11.61 | [+8.80, +13.99] | +129 / −34 | 3.15e-14 |
+| 4-bit vs 3-bit | text-to-SQL / wikisql | 533/818 | 464/818 | +8.44 | [+5.54, +10.99] | +110 / −41 | 1.79e-08 |
+| 4-bit vs 3-bit | text-to-SQL / spider | 204/818 | 115/818 | +10.88 | [+8.23, +13.07] | +118 / −29 | 6.16e-14 |
+| 4-bit vs 3-bit | HumanEval | 51/164 | 32/164 | +11.59 | [+4.46, +16.51] | +26 / −7 | 0.00132 |
+| 4-bit vs 3-bit | MBPP | 141/500 | 100/500 | +8.20 | [+4.69, +11.09] | +61 / −20 | 5.66e-06 |
+| 4-bit vs 3-bit | code (humaneval+mbpp) | 192/664 | 132/664 | +9.04 | [+5.99, +11.60] | +87 / −27 | 1.53e-08 |
+
+## Held-out loss
+
+Teacher-forced over the 999 conversations held out of the training mixture (2% of every stratum, never trained on): 90,517 assistant tokens. KL and argmax agreement compare each arm with the bf16 fine-tune token by token; NLL and token accuracy score each arm against the held-out reference text. This is the fine-tune's own training distribution, so it measures distance from the fine-tune there (for the quantized rows, what quantization did; for the base row, what fine-tuning did), not general ability.
+
+| arm | NLL (nats/token) | KL(fine-tune ‖ arm) | argmax agrees with fine-tune | token accuracy |
+|---|---:|---:|---:|---:|
+| fine-tune, bf16 (the reference) | 0.1324 | 0.0000 | 100.00% | 95.84% |
+| Kambo-v1 (base) | 0.1851 | 0.0602 | 97.24% | 94.74% |
+| DynQuant 4.25 map, encoded | 0.1485 | 0.0166 | 98.15% | 95.38% |
+| **DynQuant 4-bit, packed** | 0.1485 | 0.0166 | 98.15% | 95.38% |
+| uniform 4-bit | 0.1648 | 0.0332 | 97.25% | 94.88% |
+| permuted-signal null, 4.25 (one draw) | 0.1536 | 0.0211 | 97.81% | 95.22% |
+| DynQuant 3.25 map, encoded | 0.1912 | 0.0594 | 96.19% | 94.11% |
+| DynQuant 3-bit, packed | 0.1912 | 0.0594 | 96.19% | 94.11% |
+| uniform 3-bit | 0.3476 | 0.2118 | 91.64% | 90.24% |
+| permuted-signal null, 3.25 (one draw) | 0.2048 | 0.0718 | 95.69% | 93.75% |
+
+Paired against the controls at 4.25 bits: each row is this arm minus the control on the same tokens, summed within each conversation, with the standard error clustered by conversation, since the tokens of one conversation are not independent. A negative KL difference means this arm is closer to the fine-tune; a negative NLL difference means it puts more probability on the held-out reference text. Each pair is read on its KL z at |z| > 1.96, uncorrected (every KL result on these cards also clears a Bonferroni bar across all 4 pairs, |z| > 2.50); the NLL z is shown, not read.
+
+| against | KL difference (this arm − it) | z | conversations where this arm is closer | NLL difference | z |
+|---|---:|---:|---:|---:|---:|
+| uniform 4-bit | -0.01659 | -23.0 | 866 of 999 | -0.01626 | -15.6 |
+| permuted-signal null, 4.25 (one draw) | -0.00450 | -13.0 | 696 of 999 | -0.00509 | -7.2 |
+
+The permuted-signal null is one seeded within-role permutation (`--score-null shuffle --null-seed 0`): it moved each module's recorded score, and its measured sensitivity where it has one, to another module of the same role (190 of 205 modules moved; the other 15 drew their own place or, like the embedding, are alone in their role). No module gained or lost a measured sensitivity in the move. Its z is conditional on that draw and does not include the spread across shuffles.
+
+**Packed and encoded agree exactly on the held-out loss.** The same map scored through this repo's packed checkpoint and through bf16 encoding gives identical per-token losses and predictions, so on this loss the uniform arms (scored encoded, here and on the tasks), the null arms (scored encoded) and this arm (scored packed) differ in their bit widths and in nothing else. Task scores were not cross-checked between the two storage paths.
+
+## How the bits were allocated
+
+DynQuant gives every quantized module its own width from {2, 3, 4, 8} bits (each layer's 16 routed experts share one width per batched bank), with groups of 128 weights sharing one scale and one zero point (asymmetric). Those cost 0.25 bits per weight, so a uniform 4-bit recipe costs 4.2535 bits per parameter here, counting scales, zero points and the bf16 remainder below, and this map was asked for 4.25. It achieved **4.2479 bits** (898,002,432 bytes) against uniform 4-bit's 4.2535 (899,182,080 bytes): -0.13% bytes. The uniform figures are re-priced: DynQuant 0.5.3 charges a uniform map for 448,512 of the 499,456 bf16 remainder parameters (it leaves out the 50,944 norm weights), and here it pays for the whole remainder, as this map does.
+
+Widths come from the fine-tune itself. During training DynQuant's tracker recorded each trained module's gradient-norm variance across optimizer steps and its activation RMS and, for the 133 single-matmul modules, the channel moments from which the loss's sensitivity to quantization is computed; the allocator spends the byte budget where each byte buys the largest drop in priced damage, subject to per-role floors.
+
+| width | parameters | share |
+|---:|---:|---:|
+| 2-bit | 169,869,312 | 10.1% |
+| 3-bit | 509,607,936 | 30.1% |
+| 4-bit | 801,243,136 | 47.4% |
+| 8-bit | 209,977,344 | 12.4% |
+
+By module family (one row per projection, summed over the layers that have it; the model has 24):
+
+| family | modules | parameters | widths (share of the family's parameters) |
+|---|---:|---:|---|
+| `moe.w1` | 24 | 452,984,832 | 4b 100% |
+| `moe.w2` | 24 | 452,984,832 | 2b 21%, 3b 54%, 4b 25% |
+| `moe.w3` | 24 | 452,984,832 | 2b 17%, 3b 58%, 4b 25% |
+| `model.embed_tokens` | 1 | 155,582,464 | 8b 100% |
+| `conv.in_proj` | 18 | 56,623,104 | 4b 72%, 8b 28% |
+| `moe.sw1` | 24 | 28,311,552 | 4b 96%, 8b 4% |
+| `moe.sw2` | 24 | 28,311,552 | 4b 63%, 8b 37% |
+| `moe.sw3` | 24 | 28,311,552 | 4b 79%, 8b 21% |
+| `conv.out_proj` | 18 | 18,874,368 | 4b 67%, 8b 33% |
+| `self_attn.o_proj` | 6 | 6,291,456 | 4b 17%, 8b 83% |
+| `self_attn.q_proj` | 6 | 6,291,456 | 8b 100% |
+| `self_attn.k_proj` | 6 | 1,572,864 | 8b 100% |
+| `self_attn.v_proj` | 6 | 1,572,864 | 8b 100% |
+
+Kept in bf16 and not quantized: 499,456 parameters (0.030% of the model): `moe.router` (24 tensors, 393,216), `conv.conv` (18 tensors, 55,296), `input_layernorm` (24 tensors, 24,576), `post_attention_layernorm` (24 tensors, 24,576), `model.norm` (1 tensor, 1,024), `self_attn.k_norm` (6 tensors, 384), `self_attn.q_norm` (6 tensors, 384). These are the 24 routers, which pick each token's experts, the short-convolution kernels and the norms, all left in the compute dtype by DynQuant's classification.
+
+**72 modules holding 1,358,954,496 parameters (80.4%) were priced by a proxy, not by measurement.** DynQuant prices a module from the channel moments its hooks collect, and it collects them only for modules that are a single matmul (the tied embedding is measured through `lm_head`). Kambo stores each layer's 16 routed experts as three batched tensors (`moe.w1`, `moe.w2`, `moe.w3`) whose forward spans two matmuls and a nonlinearity, so the tracker followed those banks (`measure_expert_banks=True`; 72 recorded) and recorded their gradient statistics and activation RMS, but no moments. Those banks were priced from the tracker's plasticity score (each bank's within-role rank of log1p of its gradient-norm variance across optimizer steps; DynQuant 0.5.3's default score does not use saliency) times their size times an error curve, scaled to the measured modules' units: the allocator's fallback. The other 133 modules (331,743,232 parameters) were priced from measured moments.
+
+Every module is at or above its role's floor.
+
+**Expert banks are grouped along their output axis.** A Linear's weight is `[out, in]` and DynQuant groups along the stored last axis, the input. Kambo stores its expert banks input-major (`[experts, in, out]`), so the same rule groups each of their 128-weight blocks across 128 output channels of one input. Measured on the base model's layers 0, 12 and 23 at 4 bits, the reconstruction error of the shipped grouping relative to grouping along the input is 0.995 for `w1`, 1.017 for `w2` and 0.999 for `w3` (1.000 = no difference, above 1 = the shipped grouping is worse); the banks were not transposed.
+
+## Evaluation
+
+All scores come from `dynquant eval` (dynquant 0.5.3) with the transformers backend, bf16, the chat template, greedy decoding and every decode setting pinned identically across arms:
+
+| task | items | prompt | max new tokens | scored by |
+|---|---:|---|---:|---|
+| text-to-SQL | 2,454 | 2 solved examples as prior chat turns, then the question | 320 | execution match: the query runs against the item's database and its result set must equal the reference query's |
+| HumanEval | 164 | one user turn: complete the function, in a single code block | 1024 | the item's unit tests, pass@1 |
+| MBPP | 500 (test split) | one user turn: the task and its tests | 1024 | the item's unit tests, pass@1 |
+
+Text-to-SQL deals 818 items from each of Gretel's test split, WikiSQL's test split and Spider's 1,034-item dev set (its `validation` split), in rotation. An item is admitted only if its database holds rows and its reference query returns some, and not a single row of NULLs and zeros, so a wrong query cannot match by also returning nothing; items whose schema and rows exceed 6,000 characters are skipped. Gretel's schemas carry their own INSERTs, WikiSQL's databases are built from its real Wikipedia tables, and Spider's databases, rows included, are inlined from a mirror. MBPP's records say `shots: 3`, but the chat framing ignores exemplars (DynQuant logs that it does), so every MBPP prompt is the single turn above. Generated code runs in a sandbox (`exec/linux/py3.12/rlimits/t=8s/m=4096MB`).
+
+**Decoding is deterministic.** Kambo's experts are summed with a bf16 `index_add` whose CUDA atomics round in arrival order, and a top-2 router can turn that last bit into a different expert, so two plain runs of one checkpoint disagree on a few items. Every arm was therefore run under `torch.use_deterministic_algorithms(True)`; a repeat of 96 text-to-SQL items reproduced every prediction (EXACT). Launchers recorded across the arms above: dq_det. The base model's first, plain-launcher run is kept as a secondary row on the [bf16 card](https://huggingface.co/VikramPal/kambo-v1-sql-code), so the size of the launcher effect is on record.
+
+**Greedy is checked, not assumed.** The checkpoints were evaluated with the generation defaults inherited from the base model, which sample (below; this repo's own are greedy); the evaluation overrides them, and a check on the fine-tune confirmed that its generations are greedy: 8 of 8 generations were identical under seeds 1 and 2, and 583 of 584 generated tokens are the argmax of a teacher-forced pass over the same text; the one that is not trails it by 0.125 logits, a near-tie inside the check's 0.25-logit tolerance.
+
+Training, data and decontamination are described on the [bf16 fine-tune's card](https://huggingface.co/VikramPal/kambo-v1-sql-code).
+
+## What is not claimed
+
+- **Storage and resident memory are measured; speed is not.** This card makes no claim about decode throughput or latency against bf16.
+- **One runtime.** Loaded and scored through `transformers` with DynQuant's quantizer. Not tested with vLLM, llama.cpp, TGI or any other runtime; Kambo's architecture is custom code, so most will not load it at all.
+- **bfloat16 only.** Loading with `dtype=torch.float32` fails at the first layer (`expected scalar type Half but found Float`): the packed embedding emits its scales' dtype, which the format pins to fp16 under an fp32 model. bf16 loads and runs on GPU and CPU.
+- **The allocation is mostly proxy-priced** (see above): DynQuant's measured sensitivities cover 19.6% of the quantized parameters.
+- **Two skills.** Text-to-SQL and Python function writing, plus held-out loss on the training mixture. A quantization that holds these can lose something else.
+
+## Usage
+
+```bash
+pip install dynquant torch transformers accelerate
+```
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+import dynquant
+
+# Before from_pretrained. Without it transformers does not recognise the packed
+# format, warns, and returns a model with randomly initialised weights.
+dynquant.register_hf_quantizer()
+
+model_id = "VikramPal/kambo-v1-sql-code-DynQuant-4bit"
+tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda"
+)
+
+schema = "CREATE TABLE employees (id INTEGER, name TEXT, dept TEXT, salary INTEGER);"
+question = "Which employees in Sales earn more than 50000?"
+prompt = (
+    "Write a single SQL query that answers the question, using only the tables in the "
+    "schema. Return just the query, with no explanation.\n\n"
+    f"Schema:\n{schema}\n\nQuestion: {question}"
+)
+messages = [{"role": "user", "content": prompt}]
+inputs = tokenizer.apply_chat_template(
+    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
+).to(model.device)
+out = model.generate(**inputs, max_new_tokens=320)  # greedy: see generation_config.json
+print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
+```
+
+**This repo's `generation_config.json` is greedy, which is a change from the base model's.** Kambo-v1 ships `do_sample: true`, temperature 0.7, top_p 0.9 and top_k 2; the top_k is the MoE routing width (`top_k: 2` in config.json) carried into the generation defaults, and it restricts every sampled token to the two most likely. Every number on this card was measured greedy, so a plain `generate()` call here decodes greedily too. To sample, pass `do_sample=True` with your own `temperature`, `top_p` and `top_k`. transformers 5 still fills the unset `top_k` from config.json and warns that it "may be ignored"; greedy decoding does ignore it.
+
+### Prompt format
+
+The model was trained and evaluated on these wordings, and answers best when asked in them.
+ChatML, no system message (none is inserted when you supply none, which is how it was trained).
+
+<details><summary>Text-to-SQL</summary>
+
+```text
+Write a single SQL query that answers the question, using only the tables in the schema. Return just the query, with no explanation.
+
+Schema:
+{CREATE TABLE ... statements}
+
+Question: {question}
+```
+</details>
+
+<details><summary>Python function from a signature and docstring (HumanEval style)</summary>
+
+````text
+Complete the following Python function. Write the entire function, including the signature, inside a single ```python code block. Do not write tests, examples, or an explanation.
+
+```python
+{signature and docstring}```
+````
+</details>
+
+<details><summary>Python function from a description and tests (MBPP style)</summary>
+
+````text
+You are an expert Python programmer. Write a Python function for this task:
+
+{description}
+
+Your code must pass these tests:
+
+```python
+{assert statements}
+```
+
+Return only the function, in a single ```python code block, with no explanation.
+````
+</details>
+
+## License
+
+Released under the [Apache License 2.0](LICENSE), as the base model is; see [NOTICE](NOTICE). Training data, each under its own license: [gretelai/synthetic_text_to_sql](https://huggingface.co/datasets/gretelai/synthetic_text_to_sql) (apache-2.0), [Salesforce/wikisql](https://huggingface.co/datasets/Salesforce/wikisql) (`unknown`, as the dataset card states it), [b-mc2/sql-create-context](https://huggingface.co/datasets/b-mc2/sql-create-context) (cc-by-4.0), [nvidia/OpenCodeInstruct](https://huggingface.co/datasets/nvidia/OpenCodeInstruct) (cc-by-4.0). Evaluated on: [gretelai/synthetic_text_to_sql](https://huggingface.co/datasets/gretelai/synthetic_text_to_sql) (apache-2.0), [Salesforce/wikisql](https://huggingface.co/datasets/Salesforce/wikisql) (`unknown`, as the dataset card states it), [xlangai/spider](https://huggingface.co/datasets/xlangai/spider) (cc-by-sa-4.0), [premai-io/spider](https://huggingface.co/datasets/premai-io/spider) (no license stated on the dataset card), [openai/openai_humaneval](https://huggingface.co/datasets/openai/openai_humaneval) (mit), [google-research-datasets/mbpp](https://huggingface.co/datasets/google-research-datasets/mbpp) (cc-by-4.0).
+
+## Citation
+
+```bibtex
+@misc{kambo_v1_2026,
+  title  = {Kambo-v1: A 1.7B Hybrid Convolution-Attention Mixture-of-Experts Language Model},
+  author = {Kamboj, Vikrampal},
+  year   = {2026},
+  note   = {Apache-2.0},
+  url    = {https://huggingface.co/VikramPal/kambo-v1}
+}
+```

chat_template.jinja

@@ -0,0 +1,4 @@
+{% for message in messages %}{{ '<|im_start|>' + message['role'] + '
+' + message['content'] + '<|im_end|>' + '
+' }}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant
+' }}{% endif %}

config.json

@@ -0,0 +1,874 @@
+{
+  "architectures": [
+    "KamboForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_kambo.KamboConfig",
+    "AutoModel": "modeling_kambo.KamboModel",
+    "AutoModelForCausalLM": "modeling_kambo.KamboForCausalLM"
+  },
+  "bos_token_id": 151643,
+  "conv_kernel": 3,
+  "d_ff": 1152,
+  "dtype": "bfloat16",
+  "eos_token_id": 151645,
+  "gqa_layers": [
+    3,
+    7,
+    11,
+    15,
+    19,
+    23
+  ],
+  "head_dim": 64,
+  "hidden_size": 1024,
+  "intermediate_size": 1152,
+  "max_position_embeddings": 16384,
+  "model_type": "kambo",
+  "n_experts": 16,
+  "num_attention_heads": 16,
+  "num_experts": 16,
+  "num_experts_per_tok": 2,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 4,
+  "pad_token_id": 151643,
+  "quantization_config": {
+    "checkpoint_format": "dynquant-packed",
+    "group_size": 128,
+    "lm_head_quantized": false,
+    "modules": {
+      "model.embed_tokens": {
+        "bits": 8,
+        "out_features": 151936
+      },
+      "model.layers.0.conv.in_proj": {
+        "bits": 8,
+        "out_features": 3072
+      },
+      "model.layers.0.conv.out_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.0.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.0.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.0.moe.sw3": {
+        "bits": 8,
+        "out_features": 1152
+      },
+      "model.layers.0.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.0.moe.w2": {
+        "bits": 2,
+        "out_features": 18432
+      },
+      "model.layers.0.moe.w3": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.1.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.1.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.1.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.1.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.1.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.1.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.1.moe.w2": {
+        "bits": 2,
+        "out_features": 18432
+      },
+      "model.layers.1.moe.w3": {
+        "bits": 2,
+        "out_features": 16384
+      },
+      "model.layers.10.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.10.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.10.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.10.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.10.moe.sw3": {
+        "bits": 8,
+        "out_features": 1152
+      },
+      "model.layers.10.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.10.moe.w2": {
+        "bits": 4,
+        "out_features": 18432
+      },
+      "model.layers.10.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.11.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.11.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.11.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.11.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.11.moe.w2": {
+        "bits": 4,
+        "out_features": 18432
+      },
+      "model.layers.11.moe.w3": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.11.self_attn.k_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.11.self_attn.o_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.11.self_attn.q_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.11.self_attn.v_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.12.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.12.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.12.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.12.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.12.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.12.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.12.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.12.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.13.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.13.conv.out_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.13.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.13.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.13.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.13.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.13.moe.w2": {
+        "bits": 4,
+        "out_features": 18432
+      },
+      "model.layers.13.moe.w3": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.14.conv.in_proj": {
+        "bits": 8,
+        "out_features": 3072
+      },
+      "model.layers.14.conv.out_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.14.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.14.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.14.moe.sw3": {
+        "bits": 8,
+        "out_features": 1152
+      },
+      "model.layers.14.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.14.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.14.moe.w3": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.15.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.15.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.15.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.15.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.15.moe.w2": {
+        "bits": 4,
+        "out_features": 18432
+      },
+      "model.layers.15.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.15.self_attn.k_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.15.self_attn.o_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.15.self_attn.q_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.15.self_attn.v_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.16.conv.in_proj": {
+        "bits": 8,
+        "out_features": 3072
+      },
+      "model.layers.16.conv.out_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.16.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.16.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.16.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.16.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.16.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.16.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.17.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.17.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.17.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.17.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.17.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.17.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.17.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.17.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.18.conv.in_proj": {
+        "bits": 8,
+        "out_features": 3072
+      },
+      "model.layers.18.conv.out_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.18.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.18.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.18.moe.sw3": {
+        "bits": 8,
+        "out_features": 1152
+      },
+      "model.layers.18.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.18.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.18.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.19.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.19.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.19.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.19.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.19.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.19.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.19.self_attn.k_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.19.self_attn.o_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.19.self_attn.q_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.19.self_attn.v_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.2.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.2.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.2.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.2.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.2.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.2.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.2.moe.w2": {
+        "bits": 2,
+        "out_features": 18432
+      },
+      "model.layers.2.moe.w3": {
+        "bits": 2,
+        "out_features": 16384
+      },
+      "model.layers.20.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.20.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.20.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.20.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.20.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.20.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.20.moe.w2": {
+        "bits": 2,
+        "out_features": 18432
+      },
+      "model.layers.20.moe.w3": {
+        "bits": 2,
+        "out_features": 16384
+      },
+      "model.layers.21.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.21.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.21.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.21.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.21.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.21.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.21.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.21.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.22.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.22.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.22.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.22.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.22.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.22.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.22.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.22.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.23.moe.sw1": {
+        "bits": 8,
+        "out_features": 1152
+      },
+      "model.layers.23.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.23.moe.sw3": {
+        "bits": 8,
+        "out_features": 1152
+      },
+      "model.layers.23.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.23.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.23.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.23.self_attn.k_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.23.self_attn.o_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.23.self_attn.q_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.23.self_attn.v_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.3.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.3.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.3.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.3.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.3.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.3.moe.w3": {
+        "bits": 2,
+        "out_features": 16384
+      },
+      "model.layers.3.self_attn.k_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.3.self_attn.o_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.3.self_attn.q_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.3.self_attn.v_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.4.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.4.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.4.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.4.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.4.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.4.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.4.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.4.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.5.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.5.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.5.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.5.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.5.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.5.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.5.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.5.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.6.conv.in_proj": {
+        "bits": 8,
+        "out_features": 3072
+      },
+      "model.layers.6.conv.out_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.6.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.6.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.6.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.6.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.6.moe.w2": {
+        "bits": 2,
+        "out_features": 18432
+      },
+      "model.layers.6.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.7.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.7.moe.sw2": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.7.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.7.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.7.moe.w2": {
+        "bits": 3,
+        "out_features": 18432
+      },
+      "model.layers.7.moe.w3": {
+        "bits": 3,
+        "out_features": 16384
+      },
+      "model.layers.7.self_attn.k_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.7.self_attn.o_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.7.self_attn.q_proj": {
+        "bits": 8,
+        "out_features": 1024
+      },
+      "model.layers.7.self_attn.v_proj": {
+        "bits": 8,
+        "out_features": 256
+      },
+      "model.layers.8.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.8.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.8.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.8.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.8.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.8.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.8.moe.w2": {
+        "bits": 4,
+        "out_features": 18432
+      },
+      "model.layers.8.moe.w3": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.9.conv.in_proj": {
+        "bits": 4,
+        "out_features": 3072
+      },
+      "model.layers.9.conv.out_proj": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.9.moe.sw1": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.9.moe.sw2": {
+        "bits": 4,
+        "out_features": 1024
+      },
+      "model.layers.9.moe.sw3": {
+        "bits": 4,
+        "out_features": 1152
+      },
+      "model.layers.9.moe.w1": {
+        "bits": 4,
+        "out_features": 16384
+      },
+      "model.layers.9.moe.w2": {
+        "bits": 4,
+        "out_features": 18432
+      },
+      "model.layers.9.moe.w3": {
+        "bits": 4,
+        "out_features": 16384
+      }
+    },
+    "modules_to_not_convert": [],
+    "quant_method": "dynquant",
+    "schema_version": 1,
+    "symmetric": false,
+    "version": "0.5.3"
+  },
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 40000.0,
+  "tie_word_embeddings": true,
+  "top_k": 2,
+  "transformers_version": "5.14.1",
+  "use_cache": true,
+  "vocab_size": 151936
+}

configuration_kambo.py

@@ -0,0 +1,76 @@
+# coding=utf-8
+"""Kambo-v1 configuration."""
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class KamboConfig(PretrainedConfig):
+    """Configuration for the Kambo hybrid conv/attention MoE.
+
+    The backbone alternates two mixer types. Layers listed in ``gqa_layers``
+    (0-indexed) use grouped-query attention with RoPE and QK-norm; every other
+    layer uses a double-gated causal short convolution, which carries no
+    positional encoding and needs only a ``conv_kernel - 1`` state to decode
+    incrementally. Every layer's feed-forward is a mixture of experts:
+    ``n_experts`` routed experts at ``top_k`` plus one shared expert that runs
+    on every token.
+    """
+
+    model_type = "kambo"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=1024,
+        num_hidden_layers=24,
+        gqa_layers=(3, 7, 11, 15, 19, 23),
+        num_attention_heads=16,
+        num_key_value_heads=4,
+        head_dim=64,
+        conv_kernel=3,
+        n_experts=16,
+        top_k=2,
+        d_ff=1152,
+        max_position_embeddings=16384,
+        rope_theta=40000.0,
+        rms_norm_eps=1e-6,
+        tie_word_embeddings=True,
+        bos_token_id=151643,
+        eos_token_id=151645,
+        pad_token_id=151643,
+        use_cache=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        # JSON round-trips tuples to lists; normalise so `in` checks are stable.
+        self.gqa_layers = list(gqa_layers)
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.conv_kernel = conv_kernel
+        self.n_experts = n_experts
+        self.top_k = top_k
+        self.d_ff = d_ff
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+
+        # Aliases used by generic HF utilities and by third-party runners.
+        self.intermediate_size = d_ff
+        self.num_experts = n_experts
+        self.num_experts_per_tok = top_k
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["KamboConfig"]

dynquant_allocation.json

@@ -0,0 +1,237 @@
+{
+ "schema": "dynquant_allocation_v1",
+ "dynquant_core": "0.5.3",
+ "model": "runs/ft/model",
+ "stats": "runs/ft/stats",
+ "allocator": "sensitivity",
+ "group_size": 128,
+ "map_key": "4.25",
+ "map": {
+  "target_label": "4.25 avg bits",
+  "average_bits": 4.247889911339871,
+  "nbytes": 898002432,
+  "group_size": 128,
+  "pricing": {
+   "measured_modules": 133,
+   "proxied_modules": 72,
+   "measured_params": 331743232,
+   "proxied_params": 1358954496,
+   "proxied_share": 0.803783239010776,
+   "scale": 1.1908850493541702e-17
+  },
+  "histogram": {
+   "2": 9,
+   "3": 27,
+   "4": 119,
+   "8": 50
+  },
+  "violations": [],
+  "bits": {
+   "model.embed_tokens": 8,
+   "model.layers.0.conv.in_proj": 8,
+   "model.layers.0.conv.out_proj": 8,
+   "model.layers.0.moe.sw1": 4,
+   "model.layers.0.moe.sw2": 8,
+   "model.layers.0.moe.sw3": 8,
+   "model.layers.0.moe.w1": 4,
+   "model.layers.0.moe.w2": 2,
+   "model.layers.0.moe.w3": 4,
+   "model.layers.1.conv.in_proj": 4,
+   "model.layers.1.conv.out_proj": 4,
+   "model.layers.1.moe.sw1": 4,
+   "model.layers.1.moe.sw2": 4,
+   "model.layers.1.moe.sw3": 4,
+   "model.layers.1.moe.w1": 4,
+   "model.layers.1.moe.w2": 2,
+   "model.layers.1.moe.w3": 2,
+   "model.layers.10.conv.in_proj": 4,
+   "model.layers.10.conv.out_proj": 4,
+   "model.layers.10.moe.sw1": 4,
+   "model.layers.10.moe.sw2": 8,
+   "model.layers.10.moe.sw3": 8,
+   "model.layers.10.moe.w1": 4,
+   "model.layers.10.moe.w2": 4,
+   "model.layers.10.moe.w3": 3,
+   "model.layers.11.moe.sw1": 4,
+   "model.layers.11.moe.sw2": 4,
+   "model.layers.11.moe.sw3": 4,
+   "model.layers.11.moe.w1": 4,
+   "model.layers.11.moe.w2": 4,
+   "model.layers.11.moe.w3": 4,
+   "model.layers.11.self_attn.k_proj": 8,
+   "model.layers.11.self_attn.o_proj": 8,
+   "model.layers.11.self_attn.q_proj": 8,
+   "model.layers.11.self_attn.v_proj": 8,
+   "model.layers.12.conv.in_proj": 4,
+   "model.layers.12.conv.out_proj": 4,
+   "model.layers.12.moe.sw1": 4,
+   "model.layers.12.moe.sw2": 4,
+   "model.layers.12.moe.sw3": 4,
+   "model.layers.12.moe.w1": 4,
+   "model.layers.12.moe.w2": 3,
+   "model.layers.12.moe.w3": 3,
+   "model.layers.13.conv.in_proj": 4,
+   "model.layers.13.conv.out_proj": 8,
+   "model.layers.13.moe.sw1": 4,
+   "model.layers.13.moe.sw2": 4,
+   "model.layers.13.moe.sw3": 4,
+   "model.layers.13.moe.w1": 4,
+   "model.layers.13.moe.w2": 4,
+   "model.layers.13.moe.w3": 4,
+   "model.layers.14.conv.in_proj": 8,
+   "model.layers.14.conv.out_proj": 8,
+   "model.layers.14.moe.sw1": 4,
+   "model.layers.14.moe.sw2": 8,
+   "model.layers.14.moe.sw3": 8,
+   "model.layers.14.moe.w1": 4,
+   "model.layers.14.moe.w2": 3,
+   "model.layers.14.moe.w3": 4,
+   "model.layers.15.moe.sw1": 4,
+   "model.layers.15.moe.sw2": 4,
+   "model.layers.15.moe.sw3": 4,
+   "model.layers.15.moe.w1": 4,
+   "model.layers.15.moe.w2": 4,
+   "model.layers.15.moe.w3": 3,
+   "model.layers.15.self_attn.k_proj": 8,
+   "model.layers.15.self_attn.o_proj": 8,
+   "model.layers.15.self_attn.q_proj": 8,
+   "model.layers.15.self_attn.v_proj": 8,
+   "model.layers.16.conv.in_proj": 8,
+   "model.layers.16.conv.out_proj": 8,
+   "model.layers.16.moe.sw1": 4,
+   "model.layers.16.moe.sw2": 4,
+   "model.layers.16.moe.sw3": 4,
+   "model.layers.16.moe.w1": 4,
+   "model.layers.16.moe.w2": 3,
+   "model.layers.16.moe.w3": 3,
+   "model.layers.17.conv.in_proj": 4,
+   "model.layers.17.conv.out_proj": 4,
+   "model.layers.17.moe.sw1": 4,
+   "model.layers.17.moe.sw2": 4,
+   "model.layers.17.moe.sw3": 4,
+   "model.layers.17.moe.w1": 4,
+   "model.layers.17.moe.w2": 3,
+   "model.layers.17.moe.w3": 3,
+   "model.layers.18.conv.in_proj": 8,
+   "model.layers.18.conv.out_proj": 8,
+   "model.layers.18.moe.sw1": 4,
+   "model.layers.18.moe.sw2": 8,
+   "model.layers.18.moe.sw3": 8,
+   "model.layers.18.moe.w1": 4,
+   "model.layers.18.moe.w2": 3,
+   "model.layers.18.moe.w3": 3,
+   "model.layers.19.moe.sw1": 4,
+   "model.layers.19.moe.sw2": 8,
+   "model.layers.19.moe.sw3": 4,
+   "model.layers.19.moe.w1": 4,
+   "model.layers.19.moe.w2": 3,
+   "model.layers.19.moe.w3": 3,
+   "model.layers.19.self_attn.k_proj": 8,
+   "model.layers.19.self_attn.o_proj": 8,
+   "model.layers.19.self_attn.q_proj": 8,
+   "model.layers.19.self_attn.v_proj": 8,
+   "model.layers.2.conv.in_proj": 4,
+   "model.layers.2.conv.out_proj": 4,
+   "model.layers.2.moe.sw1": 4,
+   "model.layers.2.moe.sw2": 4,
+   "model.layers.2.moe.sw3": 4,
+   "model.layers.2.moe.w1": 4,
+   "model.layers.2.moe.w2": 2,
+   "model.layers.2.moe.w3": 2,
+   "model.layers.20.conv.in_proj": 4,
+   "model.layers.20.conv.out_proj": 4,
+   "model.layers.20.moe.sw1": 4,
+   "model.layers.20.moe.sw2": 4,
+   "model.layers.20.moe.sw3": 4,
+   "model.layers.20.moe.w1": 4,
+   "model.layers.20.moe.w2": 2,
+   "model.layers.20.moe.w3": 2,
+   "model.layers.21.conv.in_proj": 4,
+   "model.layers.21.conv.out_proj": 4,
+   "model.layers.21.moe.sw1": 4,
+   "model.layers.21.moe.sw2": 4,
+   "model.layers.21.moe.sw3": 4,
+   "model.layers.21.moe.w1": 4,
+   "model.layers.21.moe.w2": 3,
+   "model.layers.21.moe.w3": 3,
+   "model.layers.22.conv.in_proj": 4,
+   "model.layers.22.conv.out_proj": 4,
+   "model.layers.22.moe.sw1": 4,
+   "model.layers.22.moe.sw2": 8,
+   "model.layers.22.moe.sw3": 4,
+   "model.layers.22.moe.w1": 4,
+   "model.layers.22.moe.w2": 3,
+   "model.layers.22.moe.w3": 3,
+   "model.layers.23.moe.sw1": 8,
+   "model.layers.23.moe.sw2": 8,
+   "model.layers.23.moe.sw3": 8,
+   "model.layers.23.moe.w1": 4,
+   "model.layers.23.moe.w2": 3,
+   "model.layers.23.moe.w3": 3,
+   "model.layers.23.self_attn.k_proj": 8,
+   "model.layers.23.self_attn.o_proj": 4,
+   "model.layers.23.self_attn.q_proj": 8,
+   "model.layers.23.self_attn.v_proj": 8,
+   "model.layers.3.moe.sw1": 4,
+   "model.layers.3.moe.sw2": 4,
+   "model.layers.3.moe.sw3": 4,
+   "model.layers.3.moe.w1": 4,
+   "model.layers.3.moe.w2": 3,
+   "model.layers.3.moe.w3": 2,
+   "model.layers.3.self_attn.k_proj": 8,
+   "model.layers.3.self_attn.o_proj": 8,
+   "model.layers.3.self_attn.q_proj": 8,
+   "model.layers.3.self_attn.v_proj": 8,
+   "model.layers.4.conv.in_proj": 4,
+   "model.layers.4.conv.out_proj": 4,
+   "model.layers.4.moe.sw1": 4,
+   "model.layers.4.moe.sw2": 4,
+   "model.layers.4.moe.sw3": 4,
+   "model.layers.4.moe.w1": 4,
+   "model.layers.4.moe.w2": 3,
+   "model.layers.4.moe.w3": 3,
+   "model.layers.5.conv.in_proj": 4,
+   "model.layers.5.conv.out_proj": 4,
+   "model.layers.5.moe.sw1": 4,
+   "model.layers.5.moe.sw2": 4,
+   "model.layers.5.moe.sw3": 4,
+   "model.layers.5.moe.w1": 4,
+   "model.layers.5.moe.w2": 3,
+   "model.layers.5.moe.w3": 3,
+   "model.layers.6.conv.in_proj": 8,
+   "model.layers.6.conv.out_proj": 8,
+   "model.layers.6.moe.sw1": 4,
+   "model.layers.6.moe.sw2": 8,
+   "model.layers.6.moe.sw3": 4,
+   "model.layers.6.moe.w1": 4,
+   "model.layers.6.moe.w2": 2,
+   "model.layers.6.moe.w3": 3,
+   "model.layers.7.moe.sw1": 4,
+   "model.layers.7.moe.sw2": 8,
+   "model.layers.7.moe.sw3": 4,
+   "model.layers.7.moe.w1": 4,
+   "model.layers.7.moe.w2": 3,
+   "model.layers.7.moe.w3": 3,
+   "model.layers.7.self_attn.k_proj": 8,
+   "model.layers.7.self_attn.o_proj": 8,
+   "model.layers.7.self_attn.q_proj": 8,
+   "model.layers.7.self_attn.v_proj": 8,
+   "model.layers.8.conv.in_proj": 4,
+   "model.layers.8.conv.out_proj": 4,
+   "model.layers.8.moe.sw1": 4,
+   "model.layers.8.moe.sw2": 4,
+   "model.layers.8.moe.sw3": 4,
+   "model.layers.8.moe.w1": 4,
+   "model.layers.8.moe.w2": 4,
+   "model.layers.8.moe.w3": 4,
+   "model.layers.9.conv.in_proj": 4,
+   "model.layers.9.conv.out_proj": 4,
+   "model.layers.9.moe.sw1": 4,
+   "model.layers.9.moe.sw2": 4,
+   "model.layers.9.moe.sw3": 4,
+   "model.layers.9.moe.w1": 4,
+   "model.layers.9.moe.w2": 4,
+   "model.layers.9.moe.w3": 4
+  }
+ }
+}

evals/prompt_trunc.json

@@ -0,0 +1,1423 @@
+{
+ "dynquant": "0.5.3",
+ "transformers": "5.14.1",
+ "method": "prompts rebuilt through dynquant's CLI path from each record's settings and encoded with dynquant.eval.harness.encode_prompts; a prompt is cut when its length exceeds the record's max_prompt_tokens",
+ "arms": {
+  "base": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  },
+  "base-det": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  },
+  "ft": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  },
+  "dq4p": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  },
+  "u4": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  },
+  "dq3p": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  },
+  "u3": {
+   "text2sql": {
+    "items": 2454,
+    "max_prompt_tokens": 3072,
+    "truncated": 0,
+    "longest": 2878,
+    "ids_sha256": "4031706bdc463db0a2c43e89a7c0fc28e563788d514a75779fb69973a05954ae",
+    "shots": 2,
+    "shot_split": "shots",
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": {
+     "items": 43,
+     "by_source": {
+      "gretel": 22,
+      "wikisql": 21
+     },
+     "tallies": {
+      "create-context": {
+       "seen": 74577,
+       "kept": 0,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 58912,
+       "degenerate": 0,
+       "too_long": 0,
+       "failed": 1264,
+       "contaminated": 14401,
+       "errors": {
+        "OperationalError": 1037,
+        "schema": 227
+       }
+      },
+      "gretel": {
+       "seen": 40,
+       "kept": 22,
+       "not_a_query": 4,
+       "empty_result": 1,
+       "no_data": 6,
+       "degenerate": 1,
+       "too_long": 0,
+       "failed": 6,
+       "contaminated": 0,
+       "errors": {
+        "OperationalError": 6
+       }
+      },
+      "wikisql": {
+       "seen": 23,
+       "kept": 21,
+       "not_a_query": 0,
+       "empty_result": 0,
+       "no_data": 0,
+       "degenerate": 2,
+       "too_long": 0,
+       "failed": 0,
+       "contaminated": 0,
+       "errors": {}
+      }
+     },
+     "drawn": [
+      {
+       "task_id": "gretel/30621",
+       "source": "gretel"
+      },
+      {
+       "task_id": "gretel/62187",
+       "source": "gretel"
+      }
+     ]
+    },
+    "admitted_by_source": {
+     "gretel": 3055,
+     "spider": 818,
+     "wikisql": 13080
+    },
+    "tallies": {
+     "gretel": {
+      "seen": 5851,
+      "kept": 3055,
+      "not_a_query": 613,
+      "empty_result": 110,
+      "no_data": 767,
+      "degenerate": 80,
+      "too_long": 0,
+      "failed": 1226,
+      "contaminated": 0,
+      "errors": {
+       "OperationalError": 1051,
+       "schema": 167,
+       "ProgrammingError": 8
+      }
+     },
+     "spider": {
+      "seen": 1034,
+      "kept": 818,
+      "not_a_query": 0,
+      "empty_result": 168,
+      "no_data": 0,
+      "degenerate": 48,
+      "too_long": 0,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     },
+     "wikisql": {
+      "seen": 14788,
+      "kept": 13080,
+      "not_a_query": 0,
+      "empty_result": 19,
+      "no_data": 0,
+      "degenerate": 1611,
+      "too_long": 78,
+      "failed": 0,
+      "contaminated": 0,
+      "errors": {}
+     }
+    },
+    "by_source": {
+     "gretel": [
+      0,
+      818
+     ],
+     "spider": [
+      0,
+      818
+     ],
+     "wikisql": [
+      0,
+      818
+     ]
+    },
+    "longest_by_source": {
+     "gretel": 752,
+     "spider": 2542,
+     "wikisql": 2878
+    },
+    "check": {
+     "hits_reproduced": 2454,
+     "of": 2454,
+     "by_source_totals_match": true
+    },
+    "unfinished_reasoning": 0
+   },
+   "humaneval": {
+    "items": 164,
+    "max_prompt_tokens": 2048,
+    "truncated": 0,
+    "longest": 436,
+    "ids_sha256": "f615cdbdd3850559a4b9cb1926ca75c5c76b46fc18c0b3a278826b3aa4c21ad0",
+    "shots": 0,
+    "shot_split": null,
+    "prompt_style": "chat",
+    "cut_items": [],
+    "cut_lengths": [],
+    "shot_pool": null,
+    "check": {
+     "keys_match": true
+    }
+   },
+   "mbpp": {
+    "items": 500,
+    "max_prompt_tokens": 2048,
+    "truncated": 1,
+    "longest": 3745,
+    "ids_sha256": "c51d83a357256dde9927c532fe032bd0bc9b80d1ed49c0e0f0215bcfbbacf1f2",
+    "shots": 3,
+    "shot_split": "prompt",
+    "prompt_style": "chat",
+    "cut_items": [
+     "493"
+    ],
+    "cut_lengths": [
+     3745
+    ],
+    "shot_pool": {
+     "items": 10,
+     "by_source": {
+      "": 10
+     },
+     "tallies": {},
+     "drawn": [
+      {
+       "task_id": "1",
+       "source": ""
+      },
+      {
+       "task_id": "7",
+       "source": ""
+      },
+      {
+       "task_id": "10",
+       "source": ""
+      }
+     ]
+    },
+    "check": {
+     "keys_match": true
+    }
+   }
+  }
+ },
+ "problems": [],
+ "ok": true
+}

evals/results.json

@@ -0,0 +1,3031 @@
+{
+ "scores": {
+  "base": {
+   "text2sql": {
+    "correct": 1040,
+    "total": 2454,
+    "accuracy": 0.42379788101059496,
+    "seconds": 657.4,
+    "launcher": "plain",
+    "model": "/workspace/jev/kambo-v1",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      423,
+      818
+     ],
+     "wikisql": [
+      424,
+      818
+     ],
+     "spider": [
+      193,
+      818
+     ]
+    },
+    "errored": 681,
+    "exact": 506,
+    "unparseable": 5
+   },
+   "humaneval": {
+    "correct": 48,
+    "total": 164,
+    "accuracy": 0.2926829268292683,
+    "seconds": 258.4,
+    "launcher": "plain",
+    "model": "/workspace/jev/kambo-v1",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 0,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 126,
+    "total": 500,
+    "accuracy": 0.252,
+    "seconds": 850.0,
+    "launcher": "plain",
+    "model": "/workspace/jev/kambo-v1",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 2,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  },
+  "base-det": {
+   "text2sql": {
+    "correct": 1052,
+    "total": 2454,
+    "accuracy": 0.4286878565607172,
+    "seconds": 799.6,
+    "launcher": "dq_det",
+    "model": "/workspace/jev/kambo-v1",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      433,
+      818
+     ],
+     "wikisql": [
+      423,
+      818
+     ],
+     "spider": [
+      196,
+      818
+     ]
+    },
+    "errored": 682,
+    "exact": 511,
+    "unparseable": 5
+   },
+   "humaneval": {
+    "correct": 51,
+    "total": 164,
+    "accuracy": 0.31097560975609756,
+    "seconds": 283.9,
+    "launcher": "dq_det",
+    "model": "/workspace/jev/kambo-v1",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 0,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 127,
+    "total": 500,
+    "accuracy": 0.254,
+    "seconds": 882.7,
+    "launcher": "dq_det",
+    "model": "/workspace/jev/kambo-v1",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 2,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  },
+  "ft": {
+   "text2sql": {
+    "correct": 1323,
+    "total": 2454,
+    "accuracy": 0.539119804400978,
+    "seconds": 1349.7,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      492,
+      818
+     ],
+     "wikisql": [
+      621,
+      818
+     ],
+     "spider": [
+      210,
+      818
+     ]
+    },
+    "errored": 556,
+    "exact": 870,
+    "unparseable": 0
+   },
+   "humaneval": {
+    "correct": 50,
+    "total": 164,
+    "accuracy": 0.3048780487804878,
+    "seconds": 640.5,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 1,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 144,
+    "total": 500,
+    "accuracy": 0.288,
+    "seconds": 1299.0,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 1,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  },
+  "dq4p": {
+   "text2sql": {
+    "correct": 1204,
+    "total": 2454,
+    "accuracy": 0.4906275468622657,
+    "seconds": 1950.1,
+    "launcher": "dq_det",
+    "model": "runs/export/dq4",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      467,
+      818
+     ],
+     "wikisql": [
+      533,
+      818
+     ],
+     "spider": [
+      204,
+      818
+     ]
+    },
+    "errored": 624,
+    "exact": 778,
+    "unparseable": 0
+   },
+   "humaneval": {
+    "correct": 51,
+    "total": 164,
+    "accuracy": 0.31097560975609756,
+    "seconds": 693.5,
+    "launcher": "dq_det",
+    "model": "runs/export/dq4",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 0,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 141,
+    "total": 500,
+    "accuracy": 0.282,
+    "seconds": 1383.8,
+    "launcher": "dq_det",
+    "model": "runs/export/dq4",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 2,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  },
+  "dq3p": {
+   "text2sql": {
+    "correct": 951,
+    "total": 2454,
+    "accuracy": 0.38753056234718825,
+    "seconds": 2415.7,
+    "launcher": "dq_det",
+    "model": "runs/export/dq3",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      372,
+      818
+     ],
+     "wikisql": [
+      464,
+      818
+     ],
+     "spider": [
+      115,
+      818
+     ]
+    },
+    "errored": 817,
+    "exact": 576,
+    "unparseable": 0
+   },
+   "humaneval": {
+    "correct": 32,
+    "total": 164,
+    "accuracy": 0.1951219512195122,
+    "seconds": 978.3,
+    "launcher": "dq_det",
+    "model": "runs/export/dq3",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 1,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 100,
+    "total": 500,
+    "accuracy": 0.2,
+    "seconds": 1718.7,
+    "launcher": "dq_det",
+    "model": "runs/export/dq3",
+    "packed": null,
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 0,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  },
+  "u4": {
+   "text2sql": {
+    "correct": 1074,
+    "total": 2454,
+    "accuracy": 0.43765281173594134,
+    "seconds": 1481.4,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": {
+     "map": "runs/ft/maps.json",
+     "apply": "encode",
+     "group_size": 128,
+     "modules": 205,
+     "holds": "compute-dtype values; the size claim is the map's, not this model's",
+     "relative_error_median": 0.09587589744336615,
+     "relative_error_max": 0.12479219383677037
+    },
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      415,
+      818
+     ],
+     "wikisql": [
+      502,
+      818
+     ],
+     "spider": [
+      157,
+      818
+     ]
+    },
+    "errored": 710,
+    "exact": 707,
+    "unparseable": 0
+   },
+   "humaneval": {
+    "correct": 40,
+    "total": 164,
+    "accuracy": 0.24390243902439024,
+    "seconds": 653.5,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": {
+     "map": "runs/ft/maps.json",
+     "apply": "encode",
+     "group_size": 128,
+     "modules": 205,
+     "holds": "compute-dtype values; the size claim is the map's, not this model's",
+     "relative_error_median": 0.09587589744336615,
+     "relative_error_max": 0.12479219383677037
+    },
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 0,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 119,
+    "total": 500,
+    "accuracy": 0.238,
+    "seconds": 1191.6,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": {
+     "map": "runs/ft/maps.json",
+     "apply": "encode",
+     "group_size": 128,
+     "modules": 205,
+     "holds": "compute-dtype values; the size claim is the map's, not this model's",
+     "relative_error_median": 0.09587589744336615,
+     "relative_error_max": 0.12479219383677037
+    },
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 4,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  },
+  "u3": {
+   "text2sql": {
+    "correct": 597,
+    "total": 2454,
+    "accuracy": 0.24327628361858192,
+    "seconds": 2217.3,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": {
+     "map": "runs/ft/maps.json",
+     "apply": "encode",
+     "group_size": 128,
+     "modules": 205,
+     "holds": "compute-dtype values; the size claim is the map's, not this model's",
+     "relative_error_median": 0.188907566424998,
+     "relative_error_max": 0.23925855463685822
+    },
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 320,
+     "batch_size": 32,
+     "max_prompt_tokens": 3072,
+     "greedy": true
+    },
+    "by_source": {
+     "gretel": [
+      231,
+      818
+     ],
+     "wikisql": [
+      337,
+      818
+     ],
+     "spider": [
+      29,
+      818
+     ]
+    },
+    "errored": 777,
+    "exact": 406,
+    "unparseable": 0
+   },
+   "humaneval": {
+    "correct": 4,
+    "total": 164,
+    "accuracy": 0.024390243902439025,
+    "seconds": 826.3,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": {
+     "map": "runs/ft/maps.json",
+     "apply": "encode",
+     "group_size": 128,
+     "modules": 205,
+     "holds": "compute-dtype values; the size claim is the map's, not this model's",
+     "relative_error_median": 0.188907566424998,
+     "relative_error_max": 0.23925855463685822
+    },
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 0,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   },
+   "mbpp": {
+    "correct": 32,
+    "total": 500,
+    "accuracy": 0.064,
+    "seconds": 2829.1,
+    "launcher": "dq_det",
+    "model": "runs/ft/model",
+    "packed": {
+     "map": "runs/ft/maps.json",
+     "apply": "encode",
+     "group_size": 128,
+     "modules": 205,
+     "holds": "compute-dtype values; the size claim is the map's, not this model's",
+     "relative_error_median": 0.188907566424998,
+     "relative_error_max": 0.23925855463685822
+    },
+    "experts": {
+     "found": "eager",
+     "ran": "eager"
+    },
+    "decode": {
+     "max_new_tokens": 1024,
+     "batch_size": 16,
+     "max_prompt_tokens": 2048,
+     "greedy": true
+    },
+    "timeouts": 2,
+    "empty": 0,
+    "sandbox": "exec/linux/py3.12/rlimits/t=8s/m=4096MB"
+   }
+  }
+ },
+ "comparisons": [
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "text2sql",
+   "family": "primary",
+   "notes": [],
+   "n": 2454,
+   "ref_correct": 1052,
+   "arm_correct": 1323,
+   "arm_only": 387,
+   "ref_only": 116,
+   "delta_pts": 11.04319478402608,
+   "ci95_pts": [
+    9.43086582285593,
+    12.524333803864641
+   ],
+   "p": 4.256370490188364e-35,
+   "p_holm": 6.810192784301383e-34,
+   "verdict": "separated"
+  },
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 433,
+   "arm_correct": 492,
+   "arm_only": 97,
+   "ref_only": 38,
+   "delta_pts": 7.212713936430318,
+   "ci95_pts": [
+    4.446005293617086,
+    9.654277767645324
+   ],
+   "p": 3.957630638799941e-07
+  },
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 423,
+   "arm_correct": 621,
+   "arm_only": 233,
+   "ref_only": 35,
+   "delta_pts": 24.205378973105134,
+   "ci95_pts": [
+    21.169826961277405,
+    26.689991206073728
+   ],
+   "p": 4.538852113773027e-37
+  },
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 196,
+   "arm_correct": 210,
+   "arm_only": 57,
+   "ref_only": 43,
+   "delta_pts": 1.7114914425427872,
+   "ci95_pts": [
+    -0.8035760800774148,
+    4.122469031017516
+   ],
+   "p": 0.19334790449564246
+  },
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "humaneval",
+   "family": "primary",
+   "notes": [],
+   "n": 164,
+   "ref_correct": 51,
+   "arm_correct": 50,
+   "arm_only": 16,
+   "ref_only": 17,
+   "delta_pts": -0.6097560975609756,
+   "ci95_pts": [
+    -7.7283004402974536,
+    6.622356357533728
+   ],
+   "p": 1.0,
+   "p_holm": 1.0,
+   "verdict": "not separated"
+  },
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "mbpp",
+   "family": "primary",
+   "notes": [],
+   "n": 500,
+   "ref_correct": 127,
+   "arm_correct": 144,
+   "arm_only": 42,
+   "ref_only": 25,
+   "delta_pts": 3.4000000000000004,
+   "ci95_pts": [
+    0.0028071397301724587,
+    6.486292439541078
+   ],
+   "p": 0.049800114294728665,
+   "p_holm": 0.24900057147364332,
+   "verdict": "not separated"
+  },
+  {
+   "arm": "ft",
+   "ref": "base-det",
+   "label": "fine-tune vs base",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 178,
+   "arm_correct": 194,
+   "arm_only": 58,
+   "ref_only": 42,
+   "delta_pts": 2.4096385542168677,
+   "ci95_pts": [
+    -0.6891808516897325,
+    5.361881344735023
+   ],
+   "p": 0.13321061920721333
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "text2sql",
+   "family": "primary",
+   "notes": [],
+   "n": 2454,
+   "ref_correct": 1323,
+   "arm_correct": 1204,
+   "arm_only": 110,
+   "ref_only": 229,
+   "delta_pts": -4.849225753871231,
+   "ci95_pts": [
+    -6.219225399922738,
+    -3.393595321015852
+   ],
+   "p": 9.483842661366671e-11,
+   "p_holm": 1.327737972591334e-09,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 492,
+   "arm_correct": 467,
+   "arm_only": 27,
+   "ref_only": 52,
+   "delta_pts": -3.056234718826406,
+   "ci95_pts": [
+    -5.047058383073155,
+    -0.8286429356631332
+   ],
+   "p": 0.006553297670283774
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 621,
+   "arm_correct": 533,
+   "arm_only": 32,
+   "ref_only": 120,
+   "delta_pts": -10.757946210268948,
+   "ci95_pts": [
+    -13.056428050951233,
+    -8.027261194482165
+   ],
+   "p": 3.531089931549762e-13
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 210,
+   "arm_correct": 204,
+   "arm_only": 51,
+   "ref_only": 57,
+   "delta_pts": -0.7334963325183375,
+   "ci95_pts": [
+    -3.2904780054010936,
+    1.8647232504177085
+   ],
+   "p": 0.6306338345308315
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "humaneval",
+   "family": "primary",
+   "notes": [],
+   "n": 164,
+   "ref_correct": 50,
+   "arm_correct": 51,
+   "arm_only": 13,
+   "ref_only": 12,
+   "delta_pts": 0.6097560975609756,
+   "ci95_pts": [
+    -5.699480351919742,
+    6.769267997867488
+   ],
+   "p": 1.0,
+   "p_holm": 1.0,
+   "verdict": "not separated"
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "mbpp",
+   "family": "primary",
+   "notes": [],
+   "n": 500,
+   "ref_correct": 144,
+   "arm_correct": 141,
+   "arm_only": 21,
+   "ref_only": 24,
+   "delta_pts": -0.6,
+   "ci95_pts": [
+    -3.3011851860463253,
+    2.1830397104256463
+   ],
+   "p": 0.765991824244793,
+   "p_holm": 1.0,
+   "verdict": "not separated"
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ft",
+   "label": "DynQuant 4-bit vs the bf16 fine-tune",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 194,
+   "arm_correct": 192,
+   "arm_only": 34,
+   "ref_only": 36,
+   "delta_pts": -0.30120481927710846,
+   "ci95_pts": [
+    -2.8589907835760946,
+    2.2828680906024372
+   ],
+   "p": 0.9049745264594623
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "text2sql",
+   "family": "primary",
+   "notes": [],
+   "n": 2454,
+   "ref_correct": 1323,
+   "arm_correct": 951,
+   "arm_only": 91,
+   "ref_only": 463,
+   "delta_pts": -15.158924205378973,
+   "ci95_pts": [
+    -16.508782375561275,
+    -13.645063363940654
+   ],
+   "p": 5.655525565890677e-61,
+   "p_holm": 1.0179946018603218e-59,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 492,
+   "arm_correct": 372,
+   "arm_only": 29,
+   "ref_only": 149,
+   "delta_pts": -14.66992665036675,
+   "ci95_pts": [
+    -16.89005566856141,
+    -11.945222254392958
+   ],
+   "p": 1.1953898175766883e-20
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 621,
+   "arm_correct": 464,
+   "arm_only": 31,
+   "ref_only": 188,
+   "delta_pts": -19.193154034229828,
+   "ci95_pts": [
+    -21.511938091533587,
+    -16.3384243002621
+   ],
+   "p": 1.3280183862899444e-28
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 210,
+   "arm_correct": 115,
+   "arm_only": 31,
+   "ref_only": 126,
+   "delta_pts": -11.613691931540341,
+   "ci95_pts": [
+    -13.88519314216671,
+    -8.890126874773438
+   ],
+   "p": 8.674266576878856e-15
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "humaneval",
+   "family": "primary",
+   "notes": [],
+   "n": 164,
+   "ref_correct": 50,
+   "arm_correct": 32,
+   "arm_only": 8,
+   "ref_only": 26,
+   "delta_pts": -10.975609756097562,
+   "ci95_pts": [
+    -16.275973492255645,
+    -3.6608943599467514
+   ],
+   "p": 0.0029350556433200836,
+   "p_holm": 0.026415500789880753,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "mbpp",
+   "family": "primary",
+   "notes": [],
+   "n": 500,
+   "ref_correct": 144,
+   "arm_correct": 100,
+   "arm_only": 19,
+   "ref_only": 63,
+   "delta_pts": -8.799999999999999,
+   "ci95_pts": [
+    -11.623047821742047,
+    -5.3151717180651925
+   ],
+   "p": 1.1467543245820193e-06,
+   "p_holm": 1.1467543245820192e-05,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ft",
+   "label": "DynQuant 3-bit vs the bf16 fine-tune",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 194,
+   "arm_correct": 132,
+   "arm_only": 27,
+   "ref_only": 89,
+   "delta_pts": -9.33734939759036,
+   "ci95_pts": [
+    -11.902688310148866,
+    -6.2789475972142235
+   ],
+   "p": 6.4389951042590304e-09
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "text2sql",
+   "family": "primary",
+   "notes": [],
+   "n": 2454,
+   "ref_correct": 1074,
+   "arm_correct": 1204,
+   "arm_only": 273,
+   "ref_only": 143,
+   "delta_pts": 5.297473512632437,
+   "ci95_pts": [
+    3.6758347527898425,
+    6.843127024781686
+   ],
+   "p": 1.8181554605381482e-10,
+   "p_holm": 2.3636020986995925e-09,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 415,
+   "arm_correct": 467,
+   "arm_only": 98,
+   "ref_only": 46,
+   "delta_pts": 6.356968215158925,
+   "ci95_pts": [
+    3.4426276899764408,
+    9.003477473123624
+   ],
+   "p": 1.7548498975411193e-05
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 502,
+   "arm_correct": 533,
+   "arm_only": 90,
+   "ref_only": 59,
+   "delta_pts": 3.7897310513447433,
+   "ci95_pts": [
+    0.7547273919700828,
+    6.670889462397156
+   ],
+   "p": 0.013712033079849216
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 157,
+   "arm_correct": 204,
+   "arm_only": 85,
+   "ref_only": 38,
+   "delta_pts": 5.745721271393643,
+   "ci95_pts": [
+    3.050311369914509,
+    8.156830261319138
+   ],
+   "p": 2.720422673945318e-05
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "humaneval",
+   "family": "primary",
+   "notes": [],
+   "n": 164,
+   "ref_correct": 40,
+   "arm_correct": 51,
+   "arm_only": 20,
+   "ref_only": 9,
+   "delta_pts": 6.707317073170732,
+   "ci95_pts": [
+    -0.2943625127927925,
+    12.277399696767908
+   ],
+   "p": 0.06142834573984146,
+   "p_holm": 0.24900057147364332,
+   "verdict": "not separated"
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "mbpp",
+   "family": "primary",
+   "notes": [],
+   "n": 500,
+   "ref_correct": 119,
+   "arm_correct": 141,
+   "arm_only": 46,
+   "ref_only": 24,
+   "delta_pts": 4.3999999999999995,
+   "ci95_pts": [
+    0.9521129789566412,
+    7.462517028168896
+   ],
+   "p": 0.011526409654363481,
+   "p_holm": 0.07435417175292969,
+   "verdict": "not separated"
+  },
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "label": "DynQuant 4-bit vs uniform 4-bit",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 159,
+   "arm_correct": 192,
+   "arm_only": 66,
+   "ref_only": 33,
+   "delta_pts": 4.969879518072289,
+   "ci95_pts": [
+    1.9328599136565345,
+    7.700634016925539
+   ],
+   "p": 0.0011852314254676957
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "text2sql",
+   "family": "primary",
+   "notes": [],
+   "n": 2454,
+   "ref_correct": 597,
+   "arm_correct": 951,
+   "arm_only": 519,
+   "ref_only": 165,
+   "delta_pts": 14.425427872860636,
+   "ci95_pts": [
+    12.537346114538186,
+    16.18804977376616
+   ],
+   "p": 1.718266659366394e-43,
+   "p_holm": 2.9210533209228696e-42,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 231,
+   "arm_correct": 372,
+   "arm_only": 197,
+   "ref_only": 56,
+   "delta_pts": 17.237163814180928,
+   "ci95_pts": [
+    13.757502938607624,
+    20.305081268180952
+   ],
+   "p": 1.396032018692709e-19
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 337,
+   "arm_correct": 464,
+   "arm_only": 225,
+   "ref_only": 98,
+   "delta_pts": 15.52567237163814,
+   "ci95_pts": [
+    11.314417513071891,
+    19.44879278752957
+   ],
+   "p": 1.2197898902055296e-12
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 29,
+   "arm_correct": 115,
+   "arm_only": 97,
+   "ref_only": 11,
+   "delta_pts": 10.513447432762836,
+   "ci95_pts": [
+    8.583597422714222,
+    11.831127279049262
+   ],
+   "p": 2.3912424541662617e-18
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "humaneval",
+   "family": "primary",
+   "notes": [],
+   "n": 164,
+   "ref_correct": 4,
+   "arm_correct": 32,
+   "arm_only": 29,
+   "ref_only": 1,
+   "delta_pts": 17.073170731707318,
+   "ci95_pts": [
+    11.993800403655968,
+    18.261820575854983
+   ],
+   "p": 5.774199962615967e-08,
+   "p_holm": 6.92903995513916e-07,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "mbpp",
+   "family": "primary",
+   "notes": [],
+   "n": 500,
+   "ref_correct": 32,
+   "arm_correct": 100,
+   "arm_only": 80,
+   "ref_only": 12,
+   "delta_pts": 13.600000000000001,
+   "ci95_pts": [
+    10.423186041453063,
+    15.851501175375839
+   ],
+   "p": 1.715599535972955e-13,
+   "p_holm": 2.5733993039594324e-12,
+   "verdict": "separated"
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "label": "DynQuant 3-bit vs uniform 3-bit",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 36,
+   "arm_correct": 132,
+   "arm_only": 109,
+   "ref_only": 13,
+   "delta_pts": 14.457831325301203,
+   "ci95_pts": [
+    11.930927061593852,
+    16.243260478439197
+   ],
+   "p": 4.678198660220995e-20
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "text2sql",
+   "family": "secondary",
+   "notes": [
+    "launcher: dq_det vs plain"
+   ],
+   "n": 2454,
+   "ref_correct": 1052,
+   "arm_correct": 1040,
+   "arm_only": 20,
+   "ref_only": 32,
+   "delta_pts": -0.4889975550122249,
+   "ci95_pts": [
+    -1.0465835772352825,
+    0.1263560932520671
+   ],
+   "p": 0.12634707581392135
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 433,
+   "arm_correct": 423,
+   "arm_only": 3,
+   "ref_only": 13,
+   "delta_pts": -1.2224938875305624,
+   "ci95_pts": [
+    -1.7976577646711165,
+    -0.17034113943623744
+   ],
+   "p": 0.021270751953125
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 423,
+   "arm_correct": 424,
+   "arm_only": 12,
+   "ref_only": 11,
+   "delta_pts": 0.12224938875305623,
+   "ci95_pts": [
+    -1.0916396004152165,
+    1.3035422293306311
+   ],
+   "p": 1.0
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 196,
+   "arm_correct": 193,
+   "arm_only": 5,
+   "ref_only": 8,
+   "delta_pts": -0.36674816625916873,
+   "ci95_pts": [
+    -1.1487698274884883,
+    0.5855479613978435
+   ],
+   "p": 0.5810546875
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "humaneval",
+   "family": "secondary",
+   "notes": [
+    "launcher: dq_det vs plain"
+   ],
+   "n": 164,
+   "ref_correct": 51,
+   "arm_correct": 48,
+   "arm_only": 2,
+   "ref_only": 5,
+   "delta_pts": -1.8292682926829267,
+   "ci95_pts": [
+    -3.9550634594724423,
+    1.7890901876080505
+   ],
+   "p": 0.453125
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "mbpp",
+   "family": "secondary",
+   "notes": [
+    "launcher: dq_det vs plain"
+   ],
+   "n": 500,
+   "ref_correct": 127,
+   "arm_correct": 126,
+   "arm_only": 7,
+   "ref_only": 8,
+   "delta_pts": -0.2,
+   "ci95_pts": [
+    -1.7239996228841514,
+    1.4048319163356209
+   ],
+   "p": 1.0
+  },
+  {
+   "arm": "base",
+   "ref": "base-det",
+   "label": "plain vs deterministic launcher, base model",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 178,
+   "arm_correct": 174,
+   "arm_only": 9,
+   "ref_only": 13,
+   "delta_pts": -0.6024096385542169,
+   "ci95_pts": [
+    -1.9409491849586729,
+    0.9042068801041914
+   ],
+   "p": 0.5234670639038086
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "text2sql",
+   "family": "secondary",
+   "notes": [
+    "launcher: plain vs dq_det"
+   ],
+   "n": 2454,
+   "ref_correct": 1040,
+   "arm_correct": 1323,
+   "arm_only": 398,
+   "ref_only": 115,
+   "delta_pts": 11.532192339038305,
+   "ci95_pts": [
+    9.918414479279221,
+    13.011633442127323
+   ],
+   "p": 1.5898287929612967e-37
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 423,
+   "arm_correct": 492,
+   "arm_only": 105,
+   "ref_only": 36,
+   "delta_pts": 8.43520782396088,
+   "ci95_pts": [
+    5.669677726018732,
+    10.83630943172779
+   ],
+   "p": 5.07987007185982e-09
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 424,
+   "arm_correct": 621,
+   "arm_only": 232,
+   "ref_only": 35,
+   "delta_pts": 24.08312958435208,
+   "ci95_pts": [
+    21.048751622039585,
+    26.567300062048183
+   ],
+   "p": 7.898061860122764e-37
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 193,
+   "arm_correct": 210,
+   "arm_only": 61,
+   "ref_only": 44,
+   "delta_pts": 2.078239608801956,
+   "ci95_pts": [
+    -0.4963403300572413,
+    4.5325083909439865
+   ],
+   "p": 0.11799998150585828
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "humaneval",
+   "family": "secondary",
+   "notes": [
+    "launcher: plain vs dq_det"
+   ],
+   "n": 164,
+   "ref_correct": 48,
+   "arm_correct": 50,
+   "arm_only": 16,
+   "ref_only": 14,
+   "delta_pts": 1.2195121951219512,
+   "ci95_pts": [
+    -5.734564457533761,
+    7.92372881035841
+   ],
+   "p": 0.855535551905632
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "mbpp",
+   "family": "secondary",
+   "notes": [
+    "launcher: plain vs dq_det"
+   ],
+   "n": 500,
+   "ref_correct": 126,
+   "arm_correct": 144,
+   "arm_only": 40,
+   "ref_only": 22,
+   "delta_pts": 3.5999999999999996,
+   "ci95_pts": [
+    0.3314909273197259,
+    6.512125847805463
+   ],
+   "p": 0.030015864349067292
+  },
+  {
+   "arm": "ft",
+   "ref": "base",
+   "label": "fine-tune vs base, plain-launcher base",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 174,
+   "arm_correct": 194,
+   "arm_only": 56,
+   "ref_only": 36,
+   "delta_pts": 3.0120481927710845,
+   "ci95_pts": [
+    0.03792728408059196,
+    5.7866289999547815
+   ],
+   "p": 0.047011561644854
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "text2sql",
+   "family": "secondary",
+   "notes": [],
+   "n": 2454,
+   "ref_correct": 951,
+   "arm_correct": 1204,
+   "arm_only": 357,
+   "ref_only": 104,
+   "delta_pts": 10.309698451507742,
+   "ci95_pts": [
+    8.771658619176653,
+    11.713903867913626
+   ],
+   "p": 1.6368571089894648e-33
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "text2sql/gretel",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 372,
+   "arm_correct": 467,
+   "arm_only": 129,
+   "ref_only": 34,
+   "delta_pts": 11.613691931540341,
+   "ci95_pts": [
+    8.803863819612706,
+    13.988934323021963
+   ],
+   "p": 3.147438944773236e-14
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "text2sql/wikisql",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 464,
+   "arm_correct": 533,
+   "arm_only": 110,
+   "ref_only": 41,
+   "delta_pts": 8.43520782396088,
+   "ci95_pts": [
+    5.544553231117803,
+    10.986358630051567
+   ],
+   "p": 1.7850646530350223e-08
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "text2sql/spider",
+   "family": "exploratory",
+   "n": 818,
+   "ref_correct": 115,
+   "arm_correct": 204,
+   "arm_only": 118,
+   "ref_only": 29,
+   "delta_pts": 10.880195599022006,
+   "ci95_pts": [
+    8.234896309960797,
+    13.072560410224549
+   ],
+   "p": 6.155096365873754e-14
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "humaneval",
+   "family": "secondary",
+   "notes": [],
+   "n": 164,
+   "ref_correct": 32,
+   "arm_correct": 51,
+   "arm_only": 26,
+   "ref_only": 7,
+   "delta_pts": 11.585365853658537,
+   "ci95_pts": [
+    4.4638166472612095,
+    16.50787796352543
+   ],
+   "p": 0.0013187271542847157
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "mbpp",
+   "family": "secondary",
+   "notes": [],
+   "n": 500,
+   "ref_correct": 100,
+   "arm_correct": 141,
+   "arm_only": 61,
+   "ref_only": 20,
+   "delta_pts": 8.200000000000001,
+   "ci95_pts": [
+    4.689569181684579,
+    11.08700081478538
+   ],
+   "p": 5.656214894965938e-06
+  },
+  {
+   "arm": "dq4p",
+   "ref": "dq3p",
+   "label": "4-bit vs 3-bit",
+   "task": "code (humaneval+mbpp)",
+   "family": "exploratory",
+   "n": 664,
+   "ref_correct": 132,
+   "arm_correct": 192,
+   "arm_only": 87,
+   "ref_only": 27,
+   "delta_pts": 9.036144578313253,
+   "ci95_pts": [
+    5.989320681055679,
+    11.59746415825791
+   ],
+   "p": 1.5261651463770278e-08
+  }
+ ],
+ "holm_family": {
+  "declared": 18,
+  "computed": 18
+ },
+ "nll": {
+  "base": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.18506243732605415,
+    "token_acc": 0.9474242407503563,
+    "ppl": 1.20329356821513,
+    "kl": 0.06023247401703867,
+    "ref_agree": 0.9723698310814544,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": null,
+   "map_key": null,
+   "map_apply": null,
+   "model": "/workspace/jev/kambo-v1",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.17099178833328474,
+     "token_acc": 0.951563300871348,
+     "ppl": 1.1864810059951234,
+     "kl": 0.02389490271036131,
+     "ref_agree": 0.9791559883820263,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.17163155500754596,
+     "token_acc": 0.9462466487935657,
+     "ppl": 1.1872403198683597,
+     "kl": 0.01978520762472862,
+     "ref_agree": 0.9762064343163539,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.1655428069779273,
+     "token_acc": 0.9527717489600289,
+     "ppl": 1.1800334753052375,
+     "kl": 0.0434062843029191,
+     "ref_agree": 0.9809640079580394,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.1874859986788463,
+     "token_acc": 0.9486695602943952,
+     "ppl": 1.2062133607221055,
+     "kl": 0.1617937739690919,
+     "ref_agree": 0.9496131345536893,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.3505372640169584,
+     "token_acc": 0.9058461538461539,
+     "ppl": 1.4198301673702192,
+     "kl": 0.11692690633900259,
+     "ref_agree": 0.9364615384615385,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.20288218869507768,
+     "token_acc": 0.9469465648854962,
+     "ppl": 1.2249281493627555,
+     "kl": 0.17046271691595374,
+     "ref_agree": 0.9516539440203562,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": null
+  },
+  "dq3": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.191234920355092,
+    "token_acc": 0.9411049858037717,
+    "ppl": 1.2107438470493825,
+    "kl": 0.05942317319546635,
+    "ref_agree": 0.9619077079443641,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": "runs/ft/maps.json",
+   "map_key": "3.25",
+   "map_apply": "encode",
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.19619124440398966,
+     "token_acc": 0.9441312147616607,
+     "ppl": 1.2167595815455667,
+     "kl": 0.04913371599524043,
+     "ref_agree": 0.9683922774645481,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.22650612979408244,
+     "token_acc": 0.9287533512064343,
+     "ppl": 1.2542102978612026,
+     "kl": 0.06827019875144048,
+     "ref_agree": 0.953083109919571,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.19539240769655197,
+     "token_acc": 0.9374660879001627,
+     "ppl": 1.215787977455876,
+     "kl": 0.061530518396693885,
+     "ref_agree": 0.9591472237294266,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.0661412836763404,
+     "token_acc": 0.9813172296659747,
+     "ppl": 1.0683776508257397,
+     "kl": 0.03712873031590197,
+     "ref_agree": 0.9835818078882808,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.33599687224168046,
+     "token_acc": 0.904,
+     "ppl": 1.3993346480234985,
+     "kl": 0.10449129318650645,
+     "ref_agree": 0.9375384615384615,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.05812542153375446,
+     "token_acc": 0.9840966921119593,
+     "ppl": 1.0598479151256963,
+     "kl": 0.023854998211632335,
+     "ref_agree": 0.9900763358778626,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": {
+    "map": "runs/ft/maps.json",
+    "apply": "encode",
+    "group_size": 128,
+    "modules": 205,
+    "holds": "compute-dtype values; the size claim is the map's, not this model's",
+    "relative_error_median": 0.09766169726264294,
+    "relative_error_max": 0.41175080913308587
+   }
+  },
+  "dq3p": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.191234920355092,
+    "token_acc": 0.9411049858037717,
+    "ppl": 1.2107438470493825,
+    "kl": 0.05942317319546635,
+    "ref_agree": 0.9619077079443641,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": null,
+   "map_key": null,
+   "map_apply": null,
+   "model": "runs/export/dq3",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.19619124440398966,
+     "token_acc": 0.9441312147616607,
+     "ppl": 1.2167595815455667,
+     "kl": 0.04913371599524043,
+     "ref_agree": 0.9683922774645481,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.22650612979408244,
+     "token_acc": 0.9287533512064343,
+     "ppl": 1.2542102978612026,
+     "kl": 0.06827019875144048,
+     "ref_agree": 0.953083109919571,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.19539240769655197,
+     "token_acc": 0.9374660879001627,
+     "ppl": 1.215787977455876,
+     "kl": 0.061530518396693885,
+     "ref_agree": 0.9591472237294266,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.0661412836763404,
+     "token_acc": 0.9813172296659747,
+     "ppl": 1.0683776508257397,
+     "kl": 0.03712873031590197,
+     "ref_agree": 0.9835818078882808,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.33599687224168046,
+     "token_acc": 0.904,
+     "ppl": 1.3993346480234985,
+     "kl": 0.10449129318650645,
+     "ref_agree": 0.9375384615384615,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.05812542153375446,
+     "token_acc": 0.9840966921119593,
+     "ppl": 1.0598479151256963,
+     "kl": 0.023854998211632335,
+     "ref_agree": 0.9900763358778626,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": null
+  },
+  "dq4": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.14851619148530434,
+    "token_acc": 0.953820829236497,
+    "ppl": 1.160111581551982,
+    "kl": 0.016623296549019555,
+    "ref_agree": 0.9814620458035507,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": "runs/ft/maps.json",
+   "map_key": "4.25",
+   "map_apply": "encode",
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.16267120827981676,
+     "token_acc": 0.9533572526909277,
+     "ppl": 1.1766497533847577,
+     "kl": 0.013919666712902031,
+     "ref_agree": 0.985306680334871,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.18141919543851795,
+     "token_acc": 0.9407506702412869,
+     "ppl": 1.1989176547801679,
+     "kl": 0.021047137877706127,
+     "ref_agree": 0.975201072386059,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.14994969042845113,
+     "token_acc": 0.9517996020980286,
+     "ppl": 1.161775792815628,
+     "kl": 0.017496946883756036,
+     "ref_agree": 0.9799692530294809,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.036208296749001966,
+     "token_acc": 0.9903755425551991,
+     "ppl": 1.0368718010241145,
+     "kl": 0.009304864834739426,
+     "ref_agree": 0.9928288356293641,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.2619284591674805,
+     "token_acc": 0.9247692307692308,
+     "ppl": 1.2994335765441232,
+     "kl": 0.024404897270780632,
+     "ref_agree": 0.9723076923076923,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.03883703416237091,
+     "token_acc": 0.9900763358778626,
+     "ppl": 1.0396010503886395,
+     "kl": 0.005834701331977239,
+     "ref_agree": 0.9959287531806615,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": {
+    "map": "runs/ft/maps.json",
+    "apply": "encode",
+    "group_size": 128,
+    "modules": 205,
+    "holds": "compute-dtype values; the size claim is the map's, not this model's",
+    "relative_error_median": 0.0958669702797401,
+    "relative_error_max": 0.40936123283670534
+   }
+  },
+  "dq4p": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.14851619148530434,
+    "token_acc": 0.953820829236497,
+    "ppl": 1.160111581551982,
+    "kl": 0.016623296549019555,
+    "ref_agree": 0.9814620458035507,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": null,
+   "map_key": null,
+   "map_apply": null,
+   "model": "runs/export/dq4",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.16267120827981676,
+     "token_acc": 0.9533572526909277,
+     "ppl": 1.1766497533847577,
+     "kl": 0.013919666712902031,
+     "ref_agree": 0.985306680334871,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.18141919543851795,
+     "token_acc": 0.9407506702412869,
+     "ppl": 1.1989176547801679,
+     "kl": 0.021047137877706127,
+     "ref_agree": 0.975201072386059,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.14994969042845113,
+     "token_acc": 0.9517996020980286,
+     "ppl": 1.161775792815628,
+     "kl": 0.017496946883756036,
+     "ref_agree": 0.9799692530294809,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.036208296749001966,
+     "token_acc": 0.9903755425551991,
+     "ppl": 1.0368718010241145,
+     "kl": 0.009304864834739426,
+     "ref_agree": 0.9928288356293641,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.2619284591674805,
+     "token_acc": 0.9247692307692308,
+     "ppl": 1.2994335765441232,
+     "kl": 0.024404897270780632,
+     "ref_agree": 0.9723076923076923,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.03883703416237091,
+     "token_acc": 0.9900763358778626,
+     "ppl": 1.0396010503886395,
+     "kl": 0.005834701331977239,
+     "ref_agree": 0.9959287531806615,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": null
+  },
+  "ft": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.13242465121139435,
+    "token_acc": 0.9583835080703073,
+    "ppl": 1.1415929951991715,
+    "kl": 0.0,
+    "ref_agree": 1.0,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": null,
+   "map_key": null,
+   "map_apply": null,
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 2,
+    "exact": true,
+    "max_abs_token_nll_diff": 0.0,
+    "argmax_flips": 0
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.1508189712490523,
+     "token_acc": 0.9565180249444729,
+     "ppl": 1.1627861413048208,
+     "kl": 0.0,
+     "ref_agree": 1.0,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.16033798740632413,
+     "token_acc": 0.9481233243967828,
+     "ppl": 1.1739075699232349,
+     "kl": 0.0,
+     "ref_agree": 1.0,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.13211396292894997,
+     "token_acc": 0.9567733767408211,
+     "ppl": 1.1412383707239298,
+     "kl": 0.0,
+     "ref_agree": 1.0,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.02731125776352444,
+     "token_acc": 0.9933949801849405,
+     "ppl": 1.0276876287396715,
+     "kl": 0.0,
+     "ref_agree": 1.0,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.24304480596689076,
+     "token_acc": 0.929076923076923,
+     "ppl": 1.2751257560809321,
+     "kl": 0.0,
+     "ref_agree": 1.0,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.03317736669351126,
+     "token_acc": 0.9903307888040712,
+     "ppl": 1.033733872941193,
+     "kl": 0.0,
+     "ref_agree": 1.0,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": null
+  },
+  "ns3": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.20480967710607034,
+    "token_acc": 0.9375365953356828,
+    "ppl": 1.2272914610712322,
+    "kl": 0.07177002001131658,
+    "ref_agree": 0.9568589325762011,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": "runs/ft/maps.null-shuffle.json",
+   "map_key": "3.25",
+   "map_apply": "encode",
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.210966670572218,
+     "token_acc": 0.9417392790022211,
+     "ppl": 1.2348711968251334,
+     "kl": 0.062317726284401034,
+     "ref_agree": 0.964633521271143,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.2504052635811609,
+     "token_acc": 0.9212466487935657,
+     "ppl": 1.2845458908839047,
+     "kl": 0.08612487142101956,
+     "ref_agree": 0.9453083109919571,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.20834437824973018,
+     "token_acc": 0.9339166214505336,
+     "ppl": 1.231637245602335,
+     "kl": 0.07454042817381983,
+     "ref_agree": 0.9533369506239826,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.07845906748774367,
+     "token_acc": 0.9794300811473863,
+     "ppl": 1.0816190808698696,
+     "kl": 0.04964979209113991,
+     "ref_agree": 0.9826382336289866,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.34331846134479227,
+     "token_acc": 0.9027692307692308,
+     "ppl": 1.409617599170276,
+     "kl": 0.11148773608860633,
+     "ref_agree": 0.9323076923076923,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.05983740816286199,
+     "token_acc": 0.9830788804071247,
+     "ppl": 1.0616639146251674,
+     "kl": 0.025075796949903646,
+     "ref_agree": 0.9899491094147582,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": {
+    "map": "runs/ft/maps.null-shuffle.json",
+    "apply": "encode",
+    "group_size": 128,
+    "modules": 205,
+    "holds": "compute-dtype values; the size claim is the map's, not this model's",
+    "relative_error_median": 0.09766169726264294,
+    "relative_error_max": 0.41175080913308587
+   }
+  },
+  "ns4": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.15360341466703104,
+    "token_acc": 0.9521968249058188,
+    "ppl": 1.1660283653223762,
+    "kl": 0.021127189274388,
+    "ref_agree": 0.9780814653600981,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": "runs/ft/maps.null-shuffle.json",
+   "map_key": "4.25",
+   "map_apply": "encode",
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.16820271793602065,
+     "token_acc": 0.9525029899196993,
+     "ppl": 1.183176437423738,
+     "kl": 0.018502504555492193,
+     "ref_agree": 0.9826584657440629,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.1878773156183974,
+     "token_acc": 0.9400804289544236,
+     "ppl": 1.2066854648080076,
+     "kl": 0.026178602813810962,
+     "ref_agree": 0.9735254691689008,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.1558918318193932,
+     "token_acc": 0.9495840115753301,
+     "ppl": 1.1686997801236296,
+     "kl": 0.021988235157198877,
+     "ref_agree": 0.9757189365165491,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.035232734716170284,
+     "token_acc": 0.9918852613700698,
+     "ppl": 1.0358607615081898,
+     "kl": 0.00930845156272496,
+     "ref_agree": 0.9941498395923759,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.2658346262711745,
+     "token_acc": 0.922,
+     "ppl": 1.304519307625414,
+     "kl": 0.03253714658521377,
+     "ref_agree": 0.9646153846153847,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.04091334391489587,
+     "token_acc": 0.9876590330788804,
+     "ppl": 1.0417618266324207,
+     "kl": 0.009134117574190512,
+     "ref_agree": 0.9935114503816794,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": {
+    "map": "runs/ft/maps.null-shuffle.json",
+    "apply": "encode",
+    "group_size": 128,
+    "modules": 205,
+    "holds": "compute-dtype values; the size claim is the map's, not this model's",
+    "relative_error_median": 0.09593937631930032,
+    "relative_error_max": 0.40936123283670534
+   }
+  },
+  "u3": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.34758592029890434,
+    "token_acc": 0.9023608824861629,
+    "ppl": 1.4156459381116187,
+    "kl": 0.21182433540005072,
+    "ref_agree": 0.9163913960913419,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": "runs/ft/maps.json",
+   "map_key": "uniform-3",
+   "map_apply": "encode",
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.31432390628503815,
+     "token_acc": 0.9151717068170169,
+     "ppl": 1.3693332003530472,
+     "kl": 0.16750224895761034,
+     "ref_agree": 0.9301213053135144,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.4029874875782003,
+     "token_acc": 0.882171581769437,
+     "ppl": 1.4962881693577432,
+     "kl": 0.23330893886296275,
+     "ref_agree": 0.9025469168900804,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.33394813811418955,
+     "token_acc": 0.9016323024054983,
+     "ppl": 1.3964707180915412,
+     "kl": 0.19864384188609605,
+     "ref_agree": 0.9149936697413638,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.2547157235045684,
+     "token_acc": 0.9397999622570297,
+     "ppl": 1.2900948251039972,
+     "kl": 0.22134498058329602,
+     "ref_agree": 0.9422532553311945,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.6218057585496168,
+     "token_acc": 0.844,
+     "ppl": 1.8622878496388902,
+     "kl": 0.3788637918002584,
+     "ref_agree": 0.8666153846153846,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.20454398111532662,
+     "token_acc": 0.9487277353689567,
+     "ppl": 1.2269654179666425,
+     "kl": 0.1666687735565944,
+     "ref_agree": 0.9538167938931298,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": {
+    "map": "runs/ft/maps.json",
+    "apply": "encode",
+    "group_size": 128,
+    "modules": 205,
+    "holds": "compute-dtype values; the size claim is the map's, not this model's",
+    "relative_error_median": 0.188907566424998,
+    "relative_error_max": 0.23925855463685822
+   }
+  },
+  "u4": {
+   "overall": {
+    "rows": 999,
+    "tokens": 90517,
+    "nll": 0.16478117189464558,
+    "token_acc": 0.9488272921108742,
+    "ppl": 1.1791350625857513,
+    "kl": 0.0332086044061975,
+    "ref_agree": 0.9724803075665345,
+    "ref_nll": 0.13242465121139435
+   },
+   "problems": [],
+   "map": "runs/ft/maps.json",
+   "map_key": "uniform-4",
+   "map_apply": "encode",
+   "model": "runs/ft/model",
+   "repeat": {
+    "n": 1
+   },
+   "deterministic": true,
+   "dtype": "bfloat16",
+   "per_stratum": {
+    "code/opencodeinstruct/humaneval": {
+     "rows": 77,
+     "tokens": 11706,
+     "nll": 0.17306392921579408,
+     "token_acc": 0.951563300871348,
+     "ppl": 1.1889421107912812,
+     "kl": 0.02507681738564614,
+     "ref_agree": 0.9781308730565522,
+     "ref_nll": 0.1508189712490523
+    },
+    "code/opencodeinstruct/mbpp": {
+     "rows": 102,
+     "tokens": 14920,
+     "nll": 0.1976340754741638,
+     "token_acc": 0.9363941018766756,
+     "ppl": 1.2185164271857125,
+     "kl": 0.03693874788969137,
+     "ref_agree": 0.9651474530831099,
+     "ref_nll": 0.16033798740632413
+    },
+    "code/opencodeinstruct/raw": {
+     "rows": 221,
+     "tokens": 44232,
+     "nll": 0.1630951359552724,
+     "token_acc": 0.9472553807198408,
+     "ppl": 1.1771486735250913,
+     "kl": 0.03103937174116571,
+     "ref_agree": 0.9711747151383614,
+     "ref_nll": 0.13211396292894997
+    },
+    "text2sql/create-context": {
+     "rows": 200,
+     "tokens": 5299,
+     "nll": 0.08770212089594726,
+     "token_acc": 0.9786752217399509,
+     "ppl": 1.0916628900270848,
+     "kl": 0.05860063270615002,
+     "ref_agree": 0.980939799962257,
+     "ref_nll": 0.02731125776352444
+    },
+    "text2sql/gretel": {
+     "rows": 200,
+     "tokens": 6500,
+     "nll": 0.2781067943572998,
+     "token_acc": 0.9181538461538462,
+     "ppl": 1.320627225214213,
+     "kl": 0.04349825681624746,
+     "ref_agree": 0.9613846153846154,
+     "ref_nll": 0.24304480596689076
+    },
+    "text2sql/wikisql": {
+     "rows": 199,
+     "tokens": 7860,
+     "nll": 0.0578191882174737,
+     "token_acc": 0.982442748091603,
+     "ppl": 1.0595234040744779,
+     "kl": 0.024818192027286502,
+     "ref_agree": 0.9888040712468193,
+     "ref_nll": 0.03317736669351126
+    }
+   },
+   "applied": {
+    "map": "runs/ft/maps.json",
+    "apply": "encode",
+    "group_size": 128,
+    "modules": 205,
+    "holds": "compute-dtype values; the size claim is the map's, not this model's",
+    "relative_error_median": 0.09587589744336615,
+    "relative_error_max": 0.12479219383677037
+   }
+  }
+ },
+ "nll_pairs": [
+  {
+   "arm": "dq4p",
+   "ref": "u4",
+   "kl": {
+    "diff": -0.01658530785717795,
+    "se": 0.0007209748506822247,
+    "z": -23.004003317846735,
+    "p": 4.2504764811125194e-117,
+    "convs_arm_lower": 866,
+    "convs_arm_higher": 133,
+    "convs": 999
+   },
+   "nll": {
+    "diff": -0.016264980409341243,
+    "se": 0.0010409436652414187,
+    "z": -15.62522637146653,
+    "p": 4.90177671705273e-55,
+    "convs_arm_lower": 717,
+    "convs_arm_higher": 282,
+    "convs": 999
+   }
+  },
+  {
+   "arm": "dq4p",
+   "ref": "ns4",
+   "kl": {
+    "diff": -0.004503892725368443,
+    "se": 0.00034605056624119037,
+    "z": -13.01512889948378,
+    "p": 1.0037278133145185e-38,
+    "convs_arm_lower": 696,
+    "convs_arm_higher": 303,
+    "convs": 999
+   },
+   "nll": {
+    "diff": -0.005087223181726708,
+    "se": 0.0007034936093170352,
+    "z": -7.231370853056476,
+    "p": 4.781426656720506e-13,
+    "convs_arm_lower": 616,
+    "convs_arm_higher": 383,
+    "convs": 999
+   }
+  },
+  {
+   "arm": "dq3p",
+   "ref": "u3",
+   "kl": {
+    "diff": -0.15240116220458436,
+    "se": 0.0031803181947206768,
+    "z": -47.920098830856,
+    "p": 0.0,
+    "convs_arm_lower": 990,
+    "convs_arm_higher": 9,
+    "convs": 999
+   },
+   "nll": {
+    "diff": -0.15635099994381235,
+    "se": 0.003512677237772325,
+    "z": -44.51049423572069,
+    "p": 0.0,
+    "convs_arm_lower": 967,
+    "convs_arm_higher": 32,
+    "convs": 999
+   }
+  },
+  {
+   "arm": "dq3p",
+   "ref": "ns3",
+   "kl": {
+    "diff": -0.012346846815850238,
+    "se": 0.0008226269430530806,
+    "z": -15.009047442606738,
+    "p": 6.406107106180236e-51,
+    "convs_arm_lower": 687,
+    "convs_arm_higher": 312,
+    "convs": 999
+   },
+   "nll": {
+    "diff": -0.01357475675097836,
+    "se": 0.0011687353357687277,
+    "z": -11.614910866068456,
+    "p": 3.4614918124621655e-31,
+    "convs_arm_lower": 633,
+    "convs_arm_higher": 366,
+    "convs": 999
+   }
+  }
+ ],
+ "maps": {
+  "real": {
+   "path": "runs/ft/maps.json",
+   "group_size": 128,
+   "allocator": "sensitivity",
+   "keys": {
+    "4.25": {
+     "average_bits": 4.247889911339871,
+     "nbytes": 898002432,
+     "histogram": {
+      "2": 9,
+      "3": 27,
+      "4": 119,
+      "8": 50
+     },
+     "pricing": {
+      "measured_modules": 133,
+      "proxied_modules": 72,
+      "measured_params": 331743232,
+      "proxied_params": 1358954496,
+      "proxied_share": 0.803783239010776,
+      "scale": 1.1908850493541702e-17
+     },
+     "breaches": 0,
+     "breach_roles": [],
+     "modules": 205
+    },
+    "3.25": {
+     "average_bits": 3.249502780629039,
+     "nbytes": 686943744,
+     "histogram": {
+      "2": 35,
+      "3": 26,
+      "4": 117,
+      "8": 27
+     },
+     "pricing": {
+      "measured_modules": 133,
+      "proxied_modules": 72,
+      "measured_params": 331743232,
+      "proxied_params": 1358954496,
+      "proxied_share": 0.803783239010776,
+      "scale": 1.1908850493541702e-17
+     },
+     "breaches": 12,
+     "breach_roles": [
+      {
+       "role": "embedding",
+       "bits": 4,
+       "floor": 8,
+       "count": 1
+      },
+      {
+       "role": "moe.expert.gate",
+       "bits": 2,
+       "floor": 4,
+       "count": 3
+      },
+      {
+       "role": "moe.expert.gate",
+       "bits": 3,
+       "floor": 4,
+       "count": 8
+      }
+     ],
+     "modules": 205
+    },
+    "uniform-4": {
+     "average_bits": 4.2529881222886425,
+     "nbytes": 899080192,
+     "histogram": {
+      "4": 205
+     },
+     "pricing": null,
+     "breaches": 1,
+     "breach_roles": [
+      {
+       "role": "embedding",
+       "bits": 4,
+       "floor": 8,
+       "count": 1
+      }
+     ],
+     "modules": 205
+    },
+    "uniform-3": {
+     "average_bits": 3.2532834491758473,
+     "nbytes": 687742976,
+     "histogram": {
+      "3": 205
+     },
+     "pricing": null,
+     "breaches": 109,
+     "breach_roles": [
+      {
+       "role": "attn.k",
+       "bits": 3,
+       "floor": 4,
+       "count": 6
+      },
+      {
+       "role": "attn.o",
+       "bits": 3,
+       "floor": 4,
+       "count": 24
+      },
+      {
+       "role": "attn.q",
+       "bits": 3,
+       "floor": 4,
+       "count": 6
+      },
+      {
+       "role": "attn.v",
+       "bits": 3,
+       "floor": 4,
+       "count": 6
+      },
+      {
+       "role": "embedding",
+       "bits": 3,
+       "floor": 8,
+       "count": 1
+      },
+      {
+       "role": "moe.expert.gate",
+       "bits": 3,
+       "floor": 4,
+       "count": 24
+      },
+      {
+       "role": "moe.shared.gate",
+       "bits": 3,
+       "floor": 4,
+       "count": 24
+      },
+      {
+       "role": "ssm.in",
+       "bits": 3,
+       "floor": 4,
+       "count": 18
+      }
+     ],
+     "modules": 205
+    }
+   },
+   "params": {
+    "total": 1691197184,
+    "quantized": 1690697728,
+    "bf16_remainder": 499456
+   },
+   "uniform_repriced": {
+    "uniform-4": {
+     "bits": 4,
+     "nbytes": 899182080,
+     "average_bits": 4.253470090924655,
+     "recorded_nbytes": 899080192,
+     "remainder_charged": 448512,
+     "remainder": 499456
+    },
+    "uniform-3": {
+     "bits": 3,
+     "nbytes": 687844864,
+     "average_bits": 3.2537654178118594,
+     "recorded_nbytes": 687742976,
+     "remainder_charged": 448512,
+     "remainder": 499456
+    }
+   }
+  },
+  "null-shuffle": {
+   "path": "runs/ft/maps.null-shuffle.json",
+   "group_size": 128,
+   "allocator": "sensitivity+null:shuffle(seed=0)",
+   "keys": {
+    "4.25": {
+     "average_bits": 4.247889911339871,
+     "nbytes": 898002432,
+     "histogram": {
+      "2": 9,
+      "3": 27,
+      "4": 119,
+      "8": 50
+     },
+     "pricing": {
+      "measured_modules": 133,
+      "proxied_modules": 72,
+      "measured_params": 331743232,
+      "proxied_params": 1358954496,
+      "proxied_share": 0.803783239010776,
+      "scale": 1.1908850493541702e-17
+     },
+     "breaches": 0,
+     "breach_roles": [],
+     "modules": 205,
+     "moved_vs_real": 63
+    },
+    "3.25": {
+     "average_bits": 3.249502780629039,
+     "nbytes": 686943744,
+     "histogram": {
+      "2": 35,
+      "3": 26,
+      "4": 117,
+      "8": 27
+     },
+     "pricing": {
+      "measured_modules": 133,
+      "proxied_modules": 72,
+      "measured_params": 331743232,
+      "proxied_params": 1358954496,
+      "proxied_share": 0.803783239010776,
+      "scale": 1.1908850493541702e-17
+     },
+     "breaches": 12,
+     "breach_roles": [
+      {
+       "role": "embedding",
+       "bits": 4,
+       "floor": 8,
+       "count": 1
+      },
+      {
+       "role": "moe.expert.gate",
+       "bits": 2,
+       "floor": 4,
+       "count": 3
+      },
+      {
+       "role": "moe.expert.gate",
+       "bits": 3,
+       "floor": 4,
+       "count": 8
+      }
+     ],
+     "modules": 205,
+     "moved_vs_real": 52
+    }
+   }
+  }
+ },
+ "vram": [
+  {
+   "model": "runs/ft/model",
+   "resident_bytes": 3382398976,
+   "resident_gib": 3.1501,
+   "safetensors_bytes": 3382428152,
+   "safetensors_gib": 3.1501,
+   "map_predicted_bytes": null,
+   "resident_over_predicted": null,
+   "parameters": {
+    "total": 1691197184,
+    "bytes": 3382394368,
+    "buffers_bytes": 0
+   }
+  },
+  {
+   "model": "runs/export/dq4",
+   "resident_bytes": 898007040,
+   "resident_gib": 0.8363,
+   "safetensors_bytes": 898079352,
+   "safetensors_gib": 0.8364,
+   "map_predicted_bytes": 898002432,
+   "resident_over_predicted": 1.0,
+   "parameters": {
+    "total": 499456,
+    "bytes": 998912,
+    "buffers_bytes": 897003520
+   }
+  },
+  {
+   "model": "runs/export/dq3",
+   "resident_bytes": 686948352,
+   "resident_gib": 0.6398,
+   "safetensors_bytes": 687020624,
+   "safetensors_gib": 0.6398,
+   "map_predicted_bytes": 686943744,
+   "resident_over_predicted": 1.0,
+   "parameters": {
+    "total": 499456,
+    "bytes": 998912,
+    "buffers_bytes": 685944832
+   }
+  }
+ ],
+ "parity": [
+  {
+   "verdict": "EXACT",
+   "encode": "evals/nll-dq4",
+   "packed": "evals/nll-dq4p",
+   "report": {
+    "agree": "equal",
+    "kl": "equal",
+    "nll": "equal",
+    "pred": "equal",
+    "ref_nll": "equal",
+    "row": "equal",
+    "target": "equal"
+   },
+   "runs": 2
+  },
+  {
+   "verdict": "EXACT",
+   "encode": "evals/nll-dq3",
+   "packed": "evals/nll-dq3p",
+   "report": {
+    "agree": "equal",
+    "kl": "equal",
+    "nll": "equal",
+    "pred": "equal",
+    "ref_nll": "equal",
+    "row": "equal",
+    "target": "equal"
+   },
+   "runs": 2
+  }
+ ],
+ "greedy_check": [
+  "deterministic=True",
+  "MERGED {\"do_sample\": false, \"num_beams\": 1, \"temperature\": 0.7, \"top_k\": 2, \"top_p\": 0.9}",
+  "SEEDS rows identical under seeds 1 and 2: 8/8",
+  "ARGMAX 583/584 emitted tokens are the teacher-forced argmax; 1 not, 0 of them by more than 0.25 logits; gaps [0.125]",
+  "GREEDY_OK"
+ ],
+ "det_verdict": "EXACT",
+ "method": {
+  "test": "McNemar exact, two-sided, on discordant items",
+  "ci": "exact conditional 95%: Clopper-Pearson on the arm's share of the discordant items, scaled by the discordant share",
+  "delta": "arm minus reference, points",
+  "holm": "step-down within the primary family",
+  "alpha": 0.05
+ }
+}

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": false,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "max_new_tokens": 512,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 151643,
+  "transformers_version": "5.14.1",
+  "use_cache": true
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78c31c455175e4845f88dabb61ea6060995b2dcee74e82c7b307b7c9bc819db9
+size 898079352

modeling_kambo.py

@@ -0,0 +1,604 @@
+# coding=utf-8
+"""Kambo-v1: a hybrid short-convolution / grouped-query-attention MoE.
+
+The backbone is 24 layers. Six of them (3, 7, 11, 15, 19, 23) are grouped-query
+attention with RoPE and QK-norm; the other eighteen are double-gated causal
+short convolutions. Every layer's feed-forward is a mixture of experts: 16
+routed experts at top-2 plus one shared expert that sees every token.
+
+Two consequences shape this file:
+
+  * Incremental decoding needs two different caches. The attention layers need
+    the usual keys and values. The convolution layers need no keys or values at
+    all -- only the last ``conv_kernel - 1`` columns of their pre-convolution
+    signal, a few kilobytes that stay constant no matter how long the context
+    grows. ``KamboCache`` holds both, and the model tells `generate` to leave
+    cache construction alone (``_supports_default_dynamic_cache`` is False).
+
+  * The convolution carries no positional encoding, so it cannot tell a padding
+    token from a real one by position. Left-padded batches therefore zero the
+    pre-convolution signal at padded positions, which is exactly what the
+    causal left-pad does at the start of a sequence. Without that, the first
+    two real tokens of a padded row convolve against the padding and a batch of
+    two prompts does not reproduce the same two prompts run one at a time.
+"""
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import transformers
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.generation import GenerationMixin
+
+from .configuration_kambo import KamboConfig
+
+
+# ---------------------------------------------------------------------------
+# Cache
+# ---------------------------------------------------------------------------
+
+class KamboCache:
+    """Per-layer state for incremental decoding.
+
+    Deliberately not a subclass of ``transformers.Cache``: that contract assumes
+    every layer stores keys and values, and eighteen of these layers store a
+    convolution window instead. The model opts out of the default cache
+    machinery and builds this itself in ``prepare_inputs_for_generation``.
+    """
+
+    def __init__(self):
+        self.key_cache: dict = {}
+        self.value_cache: dict = {}
+        self.conv_states: dict = {}
+        self._seen = 0
+
+    def get_seq_length(self, layer_idx: int = 0) -> int:
+        return self._seen
+
+    # `generate` calls this on some paths to size a new cache.
+    def get_max_cache_shape(self):
+        return None
+
+    def get_mask_sizes(self, cache_position, layer_idx: int = 0):
+        return self._seen + cache_position.shape[0], self._seen
+
+    def update_attention(self, key, value, layer_idx: int):
+        if layer_idx in self.key_cache:
+            key = torch.cat([self.key_cache[layer_idx], key], dim=2)
+            value = torch.cat([self.value_cache[layer_idx], value], dim=2)
+        self.key_cache[layer_idx] = key
+        self.value_cache[layer_idx] = value
+        return key, value
+
+    def reorder(self, beam_idx: torch.LongTensor):
+        for d in (self.key_cache, self.value_cache, self.conv_states):
+            for i, t in d.items():
+                d[i] = t.index_select(0, beam_idx.to(t.device))
+
+    # Beam search calls this name on the cache object.
+    def reorder_cache(self, beam_idx):
+        self.reorder(beam_idx)
+
+    def batch_select_indices(self, indices):
+        self.reorder(indices)
+
+    def crop(self, max_length: int):
+        """Assisted decoding rolls the cache back when a draft is rejected.
+
+        The attention layers can be sliced, but a convolution state is a sliding
+        window that cannot be reconstructed from a shorter prefix without
+        re-running the layer. Rather than return a silently wrong state, refuse:
+        the caller sees an error instead of degraded output.
+        """
+        raise NotImplementedError(
+            "Kambo caches a convolution window that cannot be cropped. "
+            "Speculative/assisted decoding is not supported; use plain generate()."
+        )
+
+    def __len__(self):
+        return self._seen
+
+
+# ---------------------------------------------------------------------------
+# Primitives
+# ---------------------------------------------------------------------------
+
+class KamboRMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x):
+        dt = x.dtype
+        x = x.float()
+        x = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        return (x * self.weight.float()).to(dt)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+
+def _rope_cache(seq: int, head_dim: int, theta: float, device, dtype):
+    inv = 1.0 / (theta ** (torch.arange(0, head_dim, 2, device=device).float() / head_dim))
+    t = torch.arange(seq, device=device).float()
+    f = torch.outer(t, inv)
+    return torch.cos(f).to(dtype), torch.sin(f).to(dtype)
+
+
+def _apply_rope(x, cos, sin):
+    """Split-half rotary embedding.
+
+    ``cos``/``sin`` are ``head_dim // 2`` wide and are NOT duplicated to the full
+    head width. The rotation pairs channel ``i`` with channel ``i + head_dim/2``.
+    This is not the interleaved convention used by most Llama-family code; the
+    weights were trained under this one, and swapping the two produces fluent
+    output that is subtly and permanently wrong.
+    """
+    x1, x2 = x.chunk(2, dim=-1)
+    return torch.cat([x1 * cos - x2 * sin, x2 * cos + x1 * sin], dim=-1)
+
+
+class KamboShortConv(nn.Module):
+    """Double-gated causal depthwise convolution.
+
+    ``in_proj`` produces three streams; the convolution runs on ``b * v`` and its
+    output is gated again by ``c``. No positional encoding of any kind.
+    """
+
+    def __init__(self, config: KamboConfig):
+        super().__init__()
+        d, k = config.hidden_size, config.conv_kernel
+        self.k = k
+        self.in_proj = nn.Linear(d, 3 * d, bias=False)
+        self.conv = nn.Conv1d(d, d, k, groups=d, bias=False)
+        self.out_proj = nn.Linear(d, d, bias=False)
+
+    def forward(self, x, cache: Optional[KamboCache] = None, layer_idx: int = 0,
+                token_mask: Optional[torch.Tensor] = None):
+        b, c, v = self.in_proj(x).chunk(3, dim=-1)
+        g = (b * v).transpose(1, 2)                      # [B, D, T]
+
+        # Padding contributes zero, matching the zeros the causal left-pad
+        # supplies at the start of a sequence.
+        if token_mask is not None:
+            g = g * token_mask[:, None, :].to(g.dtype)
+
+        if cache is None or layer_idx not in cache.conv_states:
+            past = g.new_zeros(g.shape[0], g.shape[1], self.k - 1)
+        else:
+            past = cache.conv_states[layer_idx]
+
+        full = torch.cat([past, g], dim=-1)              # [B, D, (k-1) + T]
+        if cache is not None:
+            # Keep exactly k-1 columns regardless of T (T may be 1, or shorter
+            # than k-1 on a very short prompt).
+            cache.conv_states[layer_idx] = full[..., -(self.k - 1):].detach().clone()
+
+        y = self.conv(full).transpose(1, 2)              # [B, T, D]
+        return self.out_proj(c * y)
+
+
+class KamboAttention(nn.Module):
+    def __init__(self, config: KamboConfig, layer_idx: int):
+        super().__init__()
+        d, hd = config.hidden_size, config.head_dim
+        self.layer_idx = layer_idx
+        self.nq = config.num_attention_heads
+        self.nkv = config.num_key_value_heads
+        self.hd = hd
+        self.rep = self.nq // self.nkv
+        self.q_proj = nn.Linear(d, self.nq * hd, bias=False)
+        self.k_proj = nn.Linear(d, self.nkv * hd, bias=False)
+        self.v_proj = nn.Linear(d, self.nkv * hd, bias=False)
+        self.o_proj = nn.Linear(self.nq * hd, d, bias=False)
+        self.q_norm = KamboRMSNorm(hd, config.rms_norm_eps)
+        self.k_norm = KamboRMSNorm(hd, config.rms_norm_eps)
+
+    def forward(self, x, cos, sin, attn_bias=None, cache=None, use_causal=False):
+        B, T, _ = x.shape
+        q = self.q_proj(x).view(B, T, self.nq, self.hd).transpose(1, 2)
+        k = self.k_proj(x).view(B, T, self.nkv, self.hd).transpose(1, 2)
+        v = self.v_proj(x).view(B, T, self.nkv, self.hd).transpose(1, 2)
+
+        # QK-norm first, rotary second. The reverse order also runs.
+        q, k = self.q_norm(q), self.k_norm(k)
+        q, k = _apply_rope(q, cos, sin), _apply_rope(k, cos, sin)
+
+        if cache is not None:
+            k, v = cache.update_attention(k, v, self.layer_idx)
+
+        k = k.repeat_interleave(self.rep, dim=1)
+        v = v.repeat_interleave(self.rep, dim=1)
+
+        o = F.scaled_dot_product_attention(
+            q, k, v, attn_mask=attn_bias, is_causal=use_causal
+        )
+        return self.o_proj(o.transpose(1, 2).reshape(B, T, -1))
+
+
+class KamboMoE(nn.Module):
+    """16 routed experts at top-2, plus one shared expert on every token.
+
+    Inference is exactly dropless: tokens are sorted by expert and each expert
+    runs one GEMM over its own rows. Training used a capacity-based batched
+    path for speed, which can drop an assignment when an expert is
+    oversubscribed; at inference there is no throughput reason to accept that
+    approximation, and the loop is the path the capacity version approximates.
+    """
+
+    def __init__(self, config: KamboConfig):
+        super().__init__()
+        d, dff, E = config.hidden_size, config.d_ff, config.n_experts
+        self.E, self.k, self.d, self.dff = E, config.top_k, d, dff
+        self.router = nn.Linear(d, E, bias=False)
+        self.w1 = nn.Parameter(torch.empty(E, d, dff))
+        self.w3 = nn.Parameter(torch.empty(E, d, dff))
+        self.w2 = nn.Parameter(torch.empty(E, dff, d))
+        self.sw1 = nn.Linear(d, dff, bias=False)
+        self.sw3 = nn.Linear(d, dff, bias=False)
+        self.sw2 = nn.Linear(dff, d, bias=False)
+
+    def forward(self, x):
+        B, T, D = x.shape
+        xf = x.reshape(-1, D)
+
+        # The router runs in fp32 and must be written out explicitly: a plain
+        # module call would be demoted to bf16 under autocast, and this is the
+        # one place in the model where that changes which experts are selected.
+        dev_type = xf.device.type
+        with torch.autocast(device_type=dev_type, enabled=False):
+            logits = F.linear(xf.float(), self.router.weight.float())
+            probs = logits.softmax(-1)
+            topv, topi = probs.topk(self.k, dim=-1)
+            topv = topv / topv.sum(-1, keepdim=True)
+
+        out = self.sw2(F.silu(self.sw1(xf)) * self.sw3(xf))
+
+        flat_e = topi.reshape(-1)
+        flat_w = topv.reshape(-1).to(x.dtype)
+        order = torch.argsort(flat_e)
+        tok = torch.div(order, self.k, rounding_mode="floor")
+        counts = torch.bincount(flat_e, minlength=self.E).tolist()
+
+        xs = xf[tok]
+        ws = flat_w[order].unsqueeze(-1)
+        ys = torch.empty_like(xs)
+        s = 0
+        for e in range(self.E):
+            n = counts[e]
+            if n == 0:
+                continue
+            xe = xs[s:s + n]
+            h = F.silu(xe @ self.w1[e]) * (xe @ self.w3[e])
+            ys[s:s + n] = h @ self.w2[e]
+            s += n
+
+        out = out.index_add(0, tok, (ys * ws).to(out.dtype))
+        return out.view(B, T, D)
+
+
+class KamboDecoderLayer(nn.Module):
+    def __init__(self, config: KamboConfig, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.is_attn = layer_idx in config.gqa_layers
+        self.input_layernorm = KamboRMSNorm(config.hidden_size, config.rms_norm_eps)
+        if self.is_attn:
+            self.self_attn = KamboAttention(config, layer_idx)
+        else:
+            self.conv = KamboShortConv(config)
+        self.post_attention_layernorm = KamboRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.moe = KamboMoE(config)
+
+    def forward(self, x, cos=None, sin=None, attn_bias=None, cache=None,
+                use_causal=False, token_mask=None):
+        h = self.input_layernorm(x)
+        if self.is_attn:
+            h = self.self_attn(h, cos, sin, attn_bias=attn_bias, cache=cache,
+                               use_causal=use_causal)
+        else:
+            h = self.conv(h, cache=cache, layer_idx=self.layer_idx,
+                          token_mask=token_mask)
+        x = x + h
+        x = x + self.moe(self.post_attention_layernorm(x))
+        return x
+
+
+# ---------------------------------------------------------------------------
+# Model
+# ---------------------------------------------------------------------------
+
+class KamboPreTrainedModel(PreTrainedModel):
+    config_class = KamboConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["KamboDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_sdpa = True
+
+    def _init_weights(self, module):
+        std = 0.02
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if getattr(module, "bias", None) is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+        elif isinstance(module, KamboRMSNorm):
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, KamboMoE):
+            for p in (module.w1, module.w2, module.w3):
+                p.data.normal_(mean=0.0, std=std)
+
+
+def _build_attn_bias(attention_mask, q_len, kv_len, past_len, device, dtype):
+    """Additive [B, 1, q_len, kv_len] mask: causal AND not-padding."""
+    q_pos = torch.arange(q_len, device=device) + past_len
+    k_pos = torch.arange(kv_len, device=device)
+    allowed = (k_pos[None, :] <= q_pos[:, None])[None, None, :, :]
+
+    if attention_mask is not None:
+        pad = attention_mask[:, None, None, :].bool()
+        allowed = allowed & pad
+
+    # A row that is entirely masked would softmax over all -inf and produce
+    # NaN, which then propagates through the whole sequence. Fully padded rows
+    # exist in real batches; let such a row attend to itself and discard the
+    # result downstream rather than poisoning the batch.
+    allowed = allowed | (~allowed.any(dim=-1, keepdim=True))
+
+    bias = torch.zeros(allowed.shape, device=device, dtype=dtype)
+    return bias.masked_fill(~allowed, torch.finfo(dtype).min)
+
+
+class KamboModel(KamboPreTrainedModel):
+    def __init__(self, config: KamboConfig):
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList(
+            [KamboDecoderLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+        self.norm = KamboRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.gradient_checkpointing = False
+        self._rope = None
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _rope_for(self, position_ids, dtype, device):
+        need = int(position_ids.max().item()) + 1
+        if self._rope is None or self._rope[0].shape[0] < need or self._rope[0].device != device:
+            size = max(need, self.config.max_position_embeddings)
+            self._rope = _rope_cache(size, self.config.head_dim,
+                                     self.config.rope_theta, device, torch.float32)
+        cos, sin = self._rope
+        # [B, T, hd/2] -> [B, 1, T, hd/2] so each row uses its own positions,
+        # which is what makes left-padded batches agree with unpadded singles.
+        return (cos[position_ids].unsqueeze(1).to(dtype),
+                sin[position_ids].unsqueeze(1).to(dtype))
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[KamboCache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ):
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else True
+        output_hidden_states = bool(output_hidden_states)
+
+        if (input_ids is None) == (inputs_embeds is None):
+            raise ValueError("Pass exactly one of input_ids or inputs_embeds.")
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        x = inputs_embeds
+        B, T, _ = x.shape
+        device = x.device
+
+        if self.gradient_checkpointing and self.training:
+            use_cache = False
+        if use_cache and past_key_values is None:
+            past_key_values = KamboCache()
+        past_len = past_key_values.get_seq_length() if past_key_values is not None else 0
+        kv_len = past_len + T
+
+        if position_ids is None:
+            if attention_mask is not None:
+                # cumsum over the full mask handles left padding: the first real
+                # token gets position 0 no matter how much padding precedes it.
+                pos_full = (attention_mask.long().cumsum(-1) - 1).clamp(min=0)
+                position_ids = pos_full[:, -T:]
+            else:
+                position_ids = torch.arange(past_len, kv_len, device=device).unsqueeze(0).expand(B, T)
+
+        cos, sin = self._rope_for(position_ids, x.dtype, device)
+
+        # The fast path -- a single unpadded sequence -- is exactly what the
+        # training code ran, so parity is checked against it directly.
+        use_causal = attention_mask is None and past_len == 0 and T > 1
+        attn_bias = None
+        if not use_causal and not (attention_mask is None and T == 1 and past_len == 0):
+            attn_bias = _build_attn_bias(attention_mask, T, kv_len, past_len, device, x.dtype)
+
+        token_mask = attention_mask[:, -T:] if attention_mask is not None else None
+
+        all_hidden = [] if output_hidden_states else None
+        for layer in self.layers:
+            if all_hidden is not None:
+                all_hidden.append(x)
+            if self.gradient_checkpointing and self.training:
+                x = self._gradient_checkpointing_func(
+                    layer.__call__, x, cos, sin, attn_bias, past_key_values,
+                    use_causal, token_mask,
+                )
+            else:
+                x = layer(x, cos, sin, attn_bias=attn_bias, cache=past_key_values,
+                          use_causal=use_causal, token_mask=token_mask)
+
+        x = self.norm(x)
+        if all_hidden is not None:
+            all_hidden.append(x)
+
+        if past_key_values is not None:
+            past_key_values._seen = kv_len
+
+        if not return_dict:
+            return tuple(v for v in (x, past_key_values, all_hidden) if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=x,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=tuple(all_hidden) if all_hidden is not None else None,
+        )
+
+
+# transformers 5 expects a {tied: source} mapping here; 4.x expects a flat list
+# and raises on a dict. Both spellings mean the same thing -- lm_head shares the
+# embedding matrix -- so pick by version rather than pinning users to one.
+_TIED = ({"lm_head.weight": "model.embed_tokens.weight"}
+         if int(transformers.__version__.split(".")[0]) >= 5
+         else ["lm_head.weight"])
+
+
+class KamboForCausalLM(KamboPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = _TIED
+
+    def __init__(self, config: KamboConfig):
+        super().__init__(config)
+        self.model = KamboModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new):
+        self.lm_head = new
+
+    def get_decoder(self):
+        return self.model
+
+    # Tell `generate` not to build a Cache for us: eighteen of these layers
+    # hold a convolution window, not keys and values. Honoured identically by
+    # transformers 4.x and 5.x, both of which take this as the signal that the
+    # model supplies its own cache in prepare_inputs_for_generation.
+    def _supports_default_dynamic_cache(self) -> bool:
+        return False
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[KamboCache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs,
+    ):
+        return_dict = return_dict if return_dict is not None else True
+        # transformers renamed this argument; accept the older spelling too.
+        if "num_logits_to_keep" in kwargs:
+            logits_to_keep = kwargs.pop("num_logits_to_keep")
+
+        out = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+        )
+
+        h = out.last_hidden_state
+        if isinstance(logits_to_keep, int):
+            if logits_to_keep > 0:
+                h = h[:, -logits_to_keep:, :]
+        else:
+            h = h[:, logits_to_keep, :]
+        logits = self.lm_head(h).float()
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs
+            )
+
+        if not return_dict:
+            return tuple(v for v in (loss, logits, out.past_key_values) if v is not None)
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=out.past_key_values,
+            hidden_states=out.hidden_states,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        if use_cache and past_key_values is None:
+            past_key_values = KamboCache()
+
+        past_len = past_key_values.get_seq_length() if past_key_values is not None else 0
+        if past_len > 0:
+            input_ids = input_ids[:, past_len:]
+
+        position_ids = kwargs.get("position_ids")
+        if position_ids is None and attention_mask is not None:
+            position_ids = (attention_mask.long().cumsum(-1) - 1).clamp(min=0)
+        if position_ids is not None:
+            position_ids = position_ids[:, -input_ids.shape[1]:]
+
+        model_inputs = {
+            "input_ids": input_ids,
+            "past_key_values": past_key_values,
+            "attention_mask": attention_mask,
+            "position_ids": position_ids,
+            "use_cache": use_cache,
+        }
+        # Only the last position's logits are ever sampled; computing the full
+        # [B, T, 151936] head over a long prompt is pure waste.
+        if past_len == 0 and input_ids.shape[1] > 1:
+            model_inputs["logits_to_keep"] = 1
+        return model_inputs
+
+    def _reorder_cache(self, past_key_values, beam_idx):
+        if past_key_values is not None:
+            past_key_values.reorder(beam_idx)
+        return past_key_values
+
+
+__all__ = ["KamboConfig", "KamboModel", "KamboForCausalLM", "KamboPreTrainedModel", "KamboCache"]

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fd169731d2cbde95e10bf356d66d5997fd885dd8dbb6fb4684da3f23b2585d8
+size 11421892

tokenizer_config.json

@@ -0,0 +1,30 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 16384,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": null
+}