Release Hummingbird-V1 500M natural continuation checkpoint

Apache-2.0 release of the selected 9.59M-parameter causal LM checkpoint, including evaluation results and training provenance.

LICENSE

@@ -0,0 +1,201 @@
+                                 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 +1,18 @@
+Hummingbird-V1
+Copyright 2026 Hummingbird-V1 contributors
+
+The Hummingbird-V1 model package is licensed under the Apache License, Version 2.0.
+
+Training-data notices:
+
+- FineWeb-Edu, FineWeb-HQ, Cosmopedia v2 and FineMath material is made available
+  under the Open Data Commons Attribution License (ODC-By) v1.0.
+- DCLM baseline 1.0 material is made available under Creative Commons
+  Attribution 4.0 International (CC-BY 4.0).
+- TinyStories is made available under the Community Data License Agreement -
+  Sharing, Version 1.0.
+- Procedural arithmetic, worked tutorials and short-choice rationales in the
+  original base phase were generated by the project.
+
+See TRAINING_DATA.md and the bundled training provenance for frozen dataset
+revisions, proportions, filtering and evaluation-selection disclosures.

README.md

@@ -0,0 +1,117 @@
+---
+language:
+- en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+- causal-lm
+- base-model
+- small-language-model
+- custom_code
+- muon
+- hummingbird-v1
+datasets:
+- HuggingFaceFW/fineweb-edu
+- mlfoundations/dclm-baseline-1.0
+- epfml/FineWeb-HQ
+- HuggingFaceTB/smollm-corpus
+- HuggingFaceTB/finemath
+- roneneldan/TinyStories
+---
+
+# Hummingbird-V1
+
+Hummingbird-V1 is a 9,592,720-parameter English causal language model trained from
+scratch. This preview exports the best observed checkpoint after a 1.5B-token Muon base phase
+and a 500,170,752-token natural-corpus continuation (2,000,170,752 cumulative
+token presentations). It is a base completion/ranking model, not an instruction-tuned assistant.
+
+Released under `apache-2.0`; see `LICENSE` and `NOTICE`.
+
+## Architecture
+
+| Property | Value |
+|---|---:|
+| Parameters | 9,592,720 |
+| Layers / hidden size | 14 / 240 |
+| Attention | 6 query heads, 2 KV heads, head dimension 40 |
+| MLP | SwiGLU, intermediate size 640 |
+| Vocabulary | 4,096-token digit-aware byte-level BPE |
+| Maximum context | 2,048 tokens |
+| Training context | 512 tokens |
+| Embeddings | tied input/output |
+| Position / normalization | RoPE / RMSNorm with per-head QK normalization |
+
+The historical Python class is named `MicroLoopForDiffusionLM`, but this checkpoint was trained
+with ordinary left-to-right causal language modeling and a single layer-loop pass.
+
+## Training
+
+- Base-phase token presentations: **1,500,000,000**
+- Natural-continuation presentations: **500,170,752**
+- Cumulative token presentations: **2,000,170,752**
+- Unique packed train split: **1,091,660,174 tokens**
+- Optimizer: Muon for two-dimensional hidden weights, AdamW for embeddings and remaining weights
+- Peak learning rate: 3.0e-04
+- Effective batch: 262,144 tokens
+- Seed: 42
+
+Packed source totals across train/validation/held-out:
+
+| Source | Tokens |
+|---|---:|
+| `natural_cosmopedia_v2` | 168,663,969 |
+| `natural_dclm` | 211,970,663 |
+| `natural_finemath_4plus` | 53,640,449 |
+| `natural_fineweb_edu` | 570,339,212 |
+| `natural_fineweb_hq` | 109,226,534 |
+
+The natural continuation corpus was globally exact- and near-deduplicated and checked against
+label-free rendered Open SLM and ArithMark-3 prompt-plus-choice protection sets. Public benchmark
+results at 250M, 500M, 750M and 1B continuation checkpoints were used to select this 500M preview
+checkpoint. This creates checkpoint-selection bias and is disclosed explicitly; benchmark records
+and answers were not training examples. Official leaderboard results require independent
+verification.
+
+## Zero-shot evaluation
+
+All task values are percentages using normalized continuation accuracy where applicable.
+
+| Benchmark | Score |
+|---|---:|
+| HellaSwag | 27.56 |
+| ARC-Easy | 33.12 |
+| ARC-Challenge | 22.27 |
+| PIQA | 55.55 |
+| ArithMark-3 | 36.10 |
+| Chance-normalized Intelligence Index | **7.598** |
+
+These are self-evaluations. An official leaderboard result requires independent verification.
+
+## Usage
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "juinron/Hummingbird-V1"
+tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True).eval()
+
+inputs = tokenizer("The color of the sky is", return_tensors="pt")
+with torch.inference_mode():
+    output = model.generate(**inputs, max_new_tokens=32, do_sample=False, use_cache=False)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+Generation currently recomputes the active context because this preview does not implement a KV
+cache. Candidate-continuation likelihood scoring is its strongest intended interface.
+
+## Limitations
+
+At 9.6M parameters, Hummingbird-V1 has limited knowledge, reasoning, factuality and generation
+coherence.
+It is English-focused, not safety-aligned, and unsuitable for consequential medical, legal,
+financial or safety decisions. Web-derived and synthetic training data can contain errors,
+biases and undesirable material.

TRAINING_DATA.md

@@ -0,0 +1,45 @@
+# Hummingbird-V1 training-data notices
+
+This file documents the data lineage for the released `Hummingbird-V1` weights.
+Apache-2.0 applies to the project code and released model materials; it does not
+replace the licenses or notices attached to third-party source datasets.
+
+## Training lineage
+
+The selected checkpoint has 2,000,170,752 cumulative token presentations:
+
+- 1,500,000,000 presentations in the original Muon causal-pretraining phase;
+- 500,170,752 presentations in the natural-corpus continuation phase.
+
+The continuation used a 1,091,660,174-token packed training split. Its complete
+prepared-corpus totals across train, validation and held-out splits were:
+
+| Source | Frozen revision | Prepared tokens | License |
+|---|---|---:|---|
+| FineWeb-Edu (`sample-100BT`) | `87f09149ef4734204d70ed1d046ddc9ca3f2b8f9` | 570,339,212 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
+| DCLM baseline 1.0 | `a3b142c183aebe5af344955ae20836eb34dcf69b` | 211,970,663 | [CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
+| FineWeb-HQ | `e58199cdd52438d94405df1a4d8630cc5f13bf84` | 109,226,534 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
+| SmolLM-Corpus / Cosmopedia v2 | `3ba9d605774198c5868892d7a8deda78031a781f` | 168,663,969 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
+| FineMath (`finemath-4plus`) | `e92b25a616738fe95dc186b64dfb19f9c8525594` | 53,640,449 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
+
+The original base phase also used FineWeb-Edu, Cosmopedia v2, TinyStories, and
+project-generated procedural arithmetic, tutorial and short-choice-rationale
+text. TinyStories revision `f54c09fd23315a6f9c86f9dc80f725de7d8f9c64`
+is made available under CDLA-Sharing-1.0.
+
+## Filtering, deduplication and evaluation protection
+
+The natural corpus used quality admission filters, canonical exact-document
+deduplication, and 32-permutation MinHash/8-band LSH near-duplicate removal
+confirmed by exact Jaccard similarity at a 0.80 threshold. Documents were split
+deterministically by SHA-256 before packing.
+
+Rendered Open SLM and ArithMark-3 prompts plus choices were held in a protection
+index and excluded from training. Public benchmark results were nevertheless
+evaluated at the 250M, 500M, 750M and 1B continuation checkpoints and used to
+select the released 500M checkpoint. This checkpoint-selection bias is disclosed
+in the model card; benchmark records and answers were not training data.
+
+Exact quotas, filtering, source fields, split policy and prepared token counts
+are included in `training/corpus_contract.yaml`, `training/packed_metadata.json`
+and `training/provenance.json` in the release package.

chat_template.jinja

@@ -0,0 +1 @@
+{% for message in messages %}{{ '<' + message['role'] + '>\n' }}{{ message['content'] }}{{ '<turn_end>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<assistant>\n' }}{% endif %}

config.json

@@ -0,0 +1,81 @@
+{
+  "activation": "swiglu",
+  "architecture": "MicroLoopForDiffusionLM",
+  "architectures": [
+    "MicroLoopForDiffusionLM"
+  ],
+  "attention_implementation": "sdpa",
+  "attention_output_gate": false,
+  "attn_res_block_size": null,
+  "auto_map": {
+    "AutoConfig": "configuration_microloop.MicroLoopConfig",
+    "AutoModelForCausalLM": "modeling_microloop.MicroLoopForDiffusionLM"
+  },
+  "bos_token_id": 1,
+  "diffusion": {
+    "fallback_block_size": 16,
+    "objective": "absorbing_mask_mdlm",
+    "primary_block_size": 32,
+    "training_block_sizes": {
+      "16": 0.35,
+      "32": 0.6,
+      "64": 0.05
+    }
+  },
+  "dropout": 0.0,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "head_dimension": 40,
+  "hidden_size": 240,
+  "intermediate_size": 640,
+  "is_decoder": true,
+  "looping": {
+    "layers": [
+      4,
+      5,
+      6
+    ],
+    "maximum_serving_loops": 3,
+    "training_loop_counts": [
+      1
+    ]
+  },
+  "max_position_embeddings": 2048,
+  "model_type": "microloop_diffusion",
+  "mtp_enabled": false,
+  "normalization": "rmsnorm",
+  "num_attention_heads": 6,
+  "num_hidden_layers": 14,
+  "num_key_value_heads": 2,
+  "pad_token_id": 0,
+  "positional_encoding": "rope",
+  "qk_norm": "per_head",
+  "rms_norm_eps": 1e-05,
+  "rope_theta": 10000.0,
+  "swiglu_clamp": {
+    "enabled": true,
+    "gate_max": 10.0,
+    "linear_max": 10.0,
+    "linear_min": -10.0
+  },
+  "target_parameters": 10000000,
+  "tie_word_embeddings": true,
+  "tokenizer": {
+    "required_special_tokens": [
+      "<pad>",
+      "<bos>",
+      "<eos>",
+      "<mask>",
+      "<system>",
+      "<user>",
+      "<assistant>",
+      "<turn_end>",
+      "<doc_end>"
+    ],
+    "type": "byte_level_bpe",
+    "vocabulary_size": 4096
+  },
+  "transformers_version": "5.14.1",
+  "use_cache": false,
+  "vocab_size": 4096
+}

configuration_microloop.py

@@ -0,0 +1,180 @@
+"""Configuration for the MicroLoop-Diffusion model.
+
+The configuration is intentionally explicit.  It is the single source of truth for
+the parameter-count gate and is serializable by Hugging Face when Transformers is
+installed.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+try:  # Keep config inspection useful before optional HF integration is installed.
+    from transformers import PretrainedConfig
+except ImportError:  # pragma: no cover - exercised only in a minimal environment.
+
+    class PretrainedConfig:  # type: ignore[no-redef]
+        model_type = "microloop_diffusion"
+
+        def __init__(self, **kwargs: Any) -> None:
+            for key, value in kwargs.items():
+                setattr(self, key, value)
+
+        def to_dict(self) -> dict[str, Any]:
+            return dict(self.__dict__)
+
+
+class MicroLoopConfig(PretrainedConfig):
+    """Model, diffusion, and selective-looping configuration.
+
+    The defaults match the locked 10M specification.  Feature configuration is
+    stored on the model config for deterministic HF save/reload and is also emitted
+    separately as ``diffusion_config.json`` by the eventual release exporter.
+    """
+
+    model_type = "microloop_diffusion"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size: int = 8192,
+        hidden_size: int = 240,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 6,
+        num_key_value_heads: int = 2,
+        head_dimension: int = 40,
+        intermediate_size: int = 640,
+        activation: str = "swiglu",
+        normalization: str = "rmsnorm",
+        positional_encoding: str = "rope",
+        tie_word_embeddings: bool = True,
+        max_position_embeddings: int = 2048,
+        dropout: float = 0.0,
+        attention_implementation: str = "eager",
+        qk_norm: str = "none",
+        attention_output_gate: bool = False,
+        attn_res_block_size: int | None = None,
+        mtp_enabled: bool = False,
+        swiglu_clamp: dict[str, Any] | None = None,
+        rms_norm_eps: float = 1e-5,
+        rope_theta: float = 10000.0,
+        architecture: str = "MicroLoopForDiffusionLM",
+        target_parameters: int = 10_000_000,
+        diffusion: dict[str, Any] | None = None,
+        looping: dict[str, Any] | None = None,
+        tokenizer: dict[str, Any] | None = None,
+        **kwargs: Any,
+    ) -> None:
+        kwargs.setdefault("is_decoder", True)
+        kwargs.setdefault("is_encoder_decoder", False)
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+        self.vocab_size = int(vocab_size)
+        self.hidden_size = int(hidden_size)
+        self.num_hidden_layers = int(num_hidden_layers)
+        self.num_attention_heads = int(num_attention_heads)
+        self.num_key_value_heads = int(num_key_value_heads)
+        self.head_dimension = int(head_dimension)
+        self.intermediate_size = int(intermediate_size)
+        self.activation = activation
+        self.normalization = normalization
+        self.positional_encoding = positional_encoding
+        self.tie_word_embeddings = bool(tie_word_embeddings)
+        self.max_position_embeddings = int(max_position_embeddings)
+        self.dropout = float(dropout)
+        self.attention_implementation = str(attention_implementation)
+        self.qk_norm = str(qk_norm)
+        self.attention_output_gate = bool(attention_output_gate)
+        self.attn_res_block_size = (
+            int(attn_res_block_size) if attn_res_block_size is not None else None
+        )
+        self.mtp_enabled = bool(mtp_enabled)
+        self.swiglu_clamp = dict(swiglu_clamp or {})
+        self.rms_norm_eps = float(rms_norm_eps)
+        self.rope_theta = float(rope_theta)
+        self.architecture = architecture
+        self.target_parameters = int(target_parameters)
+        self.diffusion = dict(diffusion or {})
+        self.looping = dict(looping or {})
+        self.tokenizer = dict(tokenizer or {})
+        self.validate()
+
+    @property
+    def head_dim(self) -> int:
+        return self.head_dimension
+
+    @classmethod
+    def from_yaml(cls, path: str | Path) -> "MicroLoopConfig":
+        """Load the locked nested YAML layout used by the project configs."""
+
+        payload = yaml.safe_load(Path(path).read_text(encoding="utf-8")) or {}
+        model = dict(payload.get("model", payload))
+        model.pop("architecture", None) if model.get("architecture") is None else None
+        return cls(
+            **model,
+            diffusion=payload.get("diffusion", {}),
+            looping=payload.get("looping", {}),
+            tokenizer=payload.get("tokenizer", {}),
+        )
+
+    def validate(self) -> None:
+        """Raise a clear error for shape or locked-spec inconsistencies."""
+
+        positive = {
+            "vocab_size": self.vocab_size,
+            "hidden_size": self.hidden_size,
+            "num_hidden_layers": self.num_hidden_layers,
+            "num_attention_heads": self.num_attention_heads,
+            "num_key_value_heads": self.num_key_value_heads,
+            "head_dimension": self.head_dimension,
+            "intermediate_size": self.intermediate_size,
+            "max_position_embeddings": self.max_position_embeddings,
+        }
+        invalid = [name for name, value in positive.items() if value <= 0]
+        if invalid:
+            raise ValueError(f"Configuration values must be positive: {', '.join(invalid)}")
+        if self.hidden_size != self.num_attention_heads * self.head_dimension:
+            raise ValueError(
+                "hidden_size must equal num_attention_heads * head_dimension: "
+                f"{self.hidden_size} != {self.num_attention_heads} * {self.head_dimension}"
+            )
+        if self.num_attention_heads % self.num_key_value_heads:
+            raise ValueError("num_attention_heads must be divisible by num_key_value_heads")
+        if self.head_dimension % 2:
+            raise ValueError("RoPE requires an even head_dimension")
+        if self.dropout < 0.0 or self.dropout >= 1.0:
+            raise ValueError("dropout must be in [0, 1)")
+        if self.attention_implementation not in {"eager", "sdpa"}:
+            raise ValueError("attention_implementation must be eager or sdpa")
+        if self.qk_norm not in {"none", "per_head"}:
+            raise ValueError("qk_norm must be none or per_head")
+        if self.attn_res_block_size is not None and self.attn_res_block_size < 2:
+            raise ValueError("attn_res_block_size must be at least two when enabled")
+        if self.swiglu_clamp:
+            enabled = bool(self.swiglu_clamp.get("enabled", False))
+            if enabled:
+                linear_min = float(self.swiglu_clamp.get("linear_min", -10.0))
+                linear_max = float(self.swiglu_clamp.get("linear_max", 10.0))
+                gate_max = float(self.swiglu_clamp.get("gate_max", 10.0))
+                if linear_min >= linear_max:
+                    raise ValueError("swiglu_clamp linear_min must be below linear_max")
+                if gate_max <= 0:
+                    raise ValueError("swiglu_clamp gate_max must be positive")
+        if self.activation.lower() != "swiglu":
+            raise ValueError("M0 only implements the locked SwiGLU activation")
+        if self.normalization.lower() != "rmsnorm":
+            raise ValueError("M0 only implements the locked RMSNorm normalization")
+        if self.positional_encoding.lower() != "rope":
+            raise ValueError("M0 only implements the locked RoPE positional encoding")
+
+    def diffusion_dict(self) -> dict[str, Any]:
+        """Return a copy suitable for a standalone diffusion config artifact."""
+
+        return dict(self.diffusion)
+
+    def looping_dict(self) -> dict[str, Any]:
+        """Return a copy suitable for experiment logging."""
+
+        return dict(self.looping)

evaluation/arithmark_3.json

@@ -0,0 +1,144 @@
+{
+  "acc": 36.1,
+  "acc_norm": 36.1,
+  "categories": {
+    "elementary_school_math_continuation::addition::grades_1_2::easy": {
+      "acc": 26.5625,
+      "acc_norm": 26.5625,
+      "norm_correct": 34,
+      "raw_correct": 34,
+      "total": 128
+    },
+    "elementary_school_math_continuation::comparison::grades_2_3::medium": {
+      "acc": 15.909090909090908,
+      "acc_norm": 15.909090909090908,
+      "norm_correct": 7,
+      "raw_correct": 7,
+      "total": 44
+    },
+    "elementary_school_math_continuation::comparison_difference::grades_2_3::medium": {
+      "acc": 37.5,
+      "acc_norm": 37.5,
+      "norm_correct": 18,
+      "raw_correct": 18,
+      "total": 48
+    },
+    "elementary_school_math_continuation::data::grades_2_3::easy": {
+      "acc": 27.906976744186046,
+      "acc_norm": 27.906976744186046,
+      "norm_correct": 12,
+      "raw_correct": 12,
+      "total": 43
+    },
+    "elementary_school_math_continuation::division::grades_3_4::medium": {
+      "acc": 37.03703703703704,
+      "acc_norm": 37.03703703703704,
+      "norm_correct": 20,
+      "raw_correct": 20,
+      "total": 54
+    },
+    "elementary_school_math_continuation::fractions_counting::grades_3_4::medium": {
+      "acc": 24.0,
+      "acc_norm": 26.0,
+      "norm_correct": 13,
+      "raw_correct": 12,
+      "total": 50
+    },
+    "elementary_school_math_continuation::geometry_area::grades_4_5::medium": {
+      "acc": 53.84615384615385,
+      "acc_norm": 53.84615384615385,
+      "norm_correct": 28,
+      "raw_correct": 28,
+      "total": 52
+    },
+    "elementary_school_math_continuation::geometry_perimeter::grades_4_5::medium": {
+      "acc": 46.666666666666664,
+      "acc_norm": 46.666666666666664,
+      "norm_correct": 21,
+      "raw_correct": 21,
+      "total": 45
+    },
+    "elementary_school_math_continuation::measurement::grades_2_3::easy": {
+      "acc": 36.84210526315789,
+      "acc_norm": 36.84210526315789,
+      "norm_correct": 28,
+      "raw_correct": 28,
+      "total": 76
+    },
+    "elementary_school_math_continuation::money::grades_3_4::medium": {
+      "acc": 32.8125,
+      "acc_norm": 32.8125,
+      "norm_correct": 21,
+      "raw_correct": 21,
+      "total": 64
+    },
+    "elementary_school_math_continuation::multiplication::grades_3_4::medium": {
+      "acc": 44.5945945945946,
+      "acc_norm": 44.5945945945946,
+      "norm_correct": 33,
+      "raw_correct": 33,
+      "total": 74
+    },
+    "elementary_school_math_continuation::patterns::grades_3_4::medium": {
+      "acc": 28.30188679245283,
+      "acc_norm": 28.30188679245283,
+      "norm_correct": 15,
+      "raw_correct": 15,
+      "total": 53
+    },
+    "elementary_school_math_continuation::subtraction::grades_1_2::easy": {
+      "acc": 26.495726495726498,
+      "acc_norm": 25.64102564102564,
+      "norm_correct": 30,
+      "raw_correct": 31,
+      "total": 117
+    },
+    "elementary_school_math_continuation::time::grades_2_3::easy": {
+      "acc": 92.72727272727272,
+      "acc_norm": 92.72727272727272,
+      "norm_correct": 51,
+      "raw_correct": 51,
+      "total": 55
+    },
+    "elementary_school_math_continuation::two_step_add_subtract::grades_2_3::medium": {
+      "acc": 26.08695652173913,
+      "acc_norm": 26.08695652173913,
+      "norm_correct": 12,
+      "raw_correct": 12,
+      "total": 46
+    },
+    "elementary_school_math_continuation::two_step_addition::grades_2_3::medium": {
+      "acc": 42.10526315789473,
+      "acc_norm": 42.10526315789473,
+      "norm_correct": 8,
+      "raw_correct": 8,
+      "total": 19
+    },
+    "elementary_school_math_continuation::two_step_subtraction::grades_2_3::medium": {
+      "acc": 31.25,
+      "acc_norm": 31.25,
+      "norm_correct": 10,
+      "raw_correct": 10,
+      "total": 32
+    }
+  },
+  "frost_evaluation": {
+    "checkpoint": "D:\\llm\\frost\\artifacts\\runs\\natural20b_pilot_muon_b32\\checkpoint-tokens-0500000000",
+    "data_sha256": "bf8ab1a5193d52cdf0e05ff0b3ca226bdfcf416cb6e75562dcbe72e7e4559435",
+    "dataset": "AxiomicLabs/Arithmark-3.0",
+    "dataset_revision": "6f6e59dd9b7e2c63455f7af7f838f9ecc3d0a746",
+    "device": "cuda",
+    "evaluator_sha256": "09667e1ea05ecd8b91dbfea14a47365b555c72622afed9fc284c127ecd5e1bcd",
+    "limit": null,
+    "precision": "bfloat16",
+    "tokenizer": "D:\\llm\\frost\\artifacts\\runs\\e3_tokenizers\\tok_4k_digit"
+  },
+  "norm_correct": 361,
+  "raw_correct": 361,
+  "timing": {
+    "evaluation_seconds": 1.190516500002559,
+    "examples_per_second": 839.9715585612215,
+    "tokenization_seconds": 0.3197372000031464
+  },
+  "total": 1000
+}

evaluation/open_slm.json

@@ -0,0 +1,311 @@
+{
+  "config": {
+    "batch_size": 64,
+    "batch_sizes": [],
+    "bootstrap_iters": 100000,
+    "device": "cuda",
+    "fewshot_seed": 1234,
+    "gen_kwargs": null,
+    "limit": null,
+    "model": "MicroLoopHarnessLM",
+    "model_args": null,
+    "numpy_seed": 1234,
+    "random_seed": 0,
+    "torch_seed": 1234,
+    "use_cache": null
+  },
+  "configs": {
+    "arc_challenge": {
+      "dataset_name": "ARC-Challenge",
+      "dataset_path": "allenai/ai2_arc",
+      "description": "",
+      "doc_to_choice": "{{choices.text}}",
+      "doc_to_decontamination_query": "Question: {{question}}\nAnswer:",
+      "doc_to_target": "{{choices.label.index(answerKey)}}",
+      "doc_to_text": "Question: {{question}}\nAnswer:",
+      "fewshot_config": {
+        "doc_to_choice": "{{choices.text}}",
+        "doc_to_target": "{{choices.label.index(answerKey)}}",
+        "doc_to_text": "Question: {{question}}\nAnswer:",
+        "fewshot_delimiter": "\n\n",
+        "fewshot_indices": null,
+        "gen_prefix": null,
+        "process_docs": null,
+        "sampler": "default",
+        "samples": null,
+        "split": null,
+        "target_delimiter": " "
+      },
+      "fewshot_delimiter": "\n\n",
+      "metadata": {
+        "config_source": "D:\\llm\\frost\\.venv\\Lib\\site-packages\\lm_eval\\tasks\\arc\\arc_challenge.yaml",
+        "version": 1.0
+      },
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc"
+        },
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc_norm"
+        }
+      ],
+      "num_fewshot": 0,
+      "output_type": "multiple_choice",
+      "repeats": 1,
+      "should_decontaminate": true,
+      "target_delimiter": " ",
+      "task": "arc_challenge",
+      "test_split": "test",
+      "training_split": "train",
+      "unsafe_code": false,
+      "validation_split": "validation"
+    },
+    "arc_easy": {
+      "dataset_name": "ARC-Easy",
+      "dataset_path": "allenai/ai2_arc",
+      "description": "",
+      "doc_to_choice": "{{choices.text}}",
+      "doc_to_decontamination_query": "Question: {{question}}\nAnswer:",
+      "doc_to_target": "{{choices.label.index(answerKey)}}",
+      "doc_to_text": "Question: {{question}}\nAnswer:",
+      "fewshot_config": {
+        "doc_to_choice": "{{choices.text}}",
+        "doc_to_target": "{{choices.label.index(answerKey)}}",
+        "doc_to_text": "Question: {{question}}\nAnswer:",
+        "fewshot_delimiter": "\n\n",
+        "fewshot_indices": null,
+        "gen_prefix": null,
+        "process_docs": null,
+        "sampler": "default",
+        "samples": null,
+        "split": null,
+        "target_delimiter": " "
+      },
+      "fewshot_delimiter": "\n\n",
+      "metadata": {
+        "config_source": "D:\\llm\\frost\\.venv\\Lib\\site-packages\\lm_eval\\tasks\\arc\\arc_easy.yaml",
+        "version": 1.0
+      },
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc"
+        },
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc_norm"
+        }
+      ],
+      "num_fewshot": 0,
+      "output_type": "multiple_choice",
+      "repeats": 1,
+      "should_decontaminate": true,
+      "target_delimiter": " ",
+      "task": "arc_easy",
+      "test_split": "test",
+      "training_split": "train",
+      "unsafe_code": false,
+      "validation_split": "validation"
+    },
+    "hellaswag": {
+      "dataset_path": "Rowan/hellaswag",
+      "description": "",
+      "doc_to_choice": "choices",
+      "doc_to_target": "{{label}}",
+      "doc_to_text": "{{query}}",
+      "fewshot_config": {
+        "doc_to_choice": "choices",
+        "doc_to_target": "{{label}}",
+        "doc_to_text": "{{query}}",
+        "fewshot_delimiter": "\n\n",
+        "fewshot_indices": null,
+        "gen_prefix": null,
+        "process_docs": "<callable function>",
+        "sampler": "default",
+        "samples": null,
+        "split": null,
+        "target_delimiter": " "
+      },
+      "fewshot_delimiter": "\n\n",
+      "metadata": {
+        "config_source": "D:\\llm\\frost\\.venv\\Lib\\site-packages\\lm_eval\\tasks\\hellaswag\\hellaswag.yaml",
+        "version": 1.0
+      },
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc"
+        },
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc_norm"
+        }
+      ],
+      "num_fewshot": 0,
+      "output_type": "multiple_choice",
+      "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n    def _process_doc(doc):\n        ctx = doc[\"ctx_a\"] + \" \" + doc[\"ctx_b\"].capitalize()\n        out_doc = {\n            \"query\": preprocess(doc[\"activity_label\"] + \": \" + ctx),\n            \"choices\": [preprocess(ending) for ending in doc[\"endings\"]],\n            \"gold\": int(doc[\"label\"]),\n        }\n        return out_doc\n\n    return dataset.map(_process_doc)\n",
+      "repeats": 1,
+      "should_decontaminate": false,
+      "target_delimiter": " ",
+      "task": "hellaswag",
+      "training_split": "train",
+      "unsafe_code": false,
+      "validation_split": "validation"
+    },
+    "piqa": {
+      "dataset_path": "baber/piqa",
+      "description": "",
+      "doc_to_choice": "{{[sol1, sol2]}}",
+      "doc_to_decontamination_query": "goal",
+      "doc_to_target": "label",
+      "doc_to_text": "Question: {{goal}}\nAnswer:",
+      "fewshot_config": {
+        "doc_to_choice": "{{[sol1, sol2]}}",
+        "doc_to_target": "label",
+        "doc_to_text": "Question: {{goal}}\nAnswer:",
+        "fewshot_delimiter": "\n\n",
+        "fewshot_indices": null,
+        "gen_prefix": null,
+        "process_docs": null,
+        "sampler": "default",
+        "samples": null,
+        "split": null,
+        "target_delimiter": " "
+      },
+      "fewshot_delimiter": "\n\n",
+      "metadata": {
+        "config_source": "D:\\llm\\frost\\.venv\\Lib\\site-packages\\lm_eval\\tasks\\piqa\\piqa.yaml",
+        "version": 1.0
+      },
+      "metric_list": [
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc"
+        },
+        {
+          "aggregation": "mean",
+          "higher_is_better": true,
+          "metric": "acc_norm"
+        }
+      ],
+      "num_fewshot": 0,
+      "output_type": "multiple_choice",
+      "repeats": 1,
+      "should_decontaminate": true,
+      "target_delimiter": " ",
+      "task": "piqa",
+      "training_split": "train",
+      "unsafe_code": false,
+      "validation_split": "validation"
+    }
+  },
+  "date": 1786715056.057948,
+  "frost_evaluation": {
+    "batch_size": 64,
+    "checkpoint": "D:\\llm\\frost\\artifacts\\runs\\natural20b_pilot_muon_b32\\checkpoint-tokens-0500000000",
+    "device": "cuda",
+    "num_fewshot": 0,
+    "tokenizer": "D:\\llm\\frost\\artifacts\\runs\\e3_tokenizers\\tok_4k_digit"
+  },
+  "git_hash": "9ac6cbb",
+  "group_subtasks": {},
+  "higher_is_better": {
+    "arc_challenge": {
+      "acc": true,
+      "acc_norm": true
+    },
+    "arc_easy": {
+      "acc": true,
+      "acc_norm": true
+    },
+    "hellaswag": {
+      "acc": true,
+      "acc_norm": true
+    },
+    "piqa": {
+      "acc": true,
+      "acc_norm": true
+    }
+  },
+  "lm_eval_version": "0.4.12",
+  "n-samples": {
+    "arc_challenge": {
+      "effective": 1172,
+      "original": 1172
+    },
+    "arc_easy": {
+      "effective": 2376,
+      "original": 2376
+    },
+    "hellaswag": {
+      "effective": 10042,
+      "original": 10042
+    },
+    "piqa": {
+      "effective": 1838,
+      "original": 1838
+    }
+  },
+  "n-shot": {
+    "arc_challenge": 0,
+    "arc_easy": 0,
+    "hellaswag": 0,
+    "piqa": 0
+  },
+  "pretty_env_info": "the JSON object must be str, bytes or bytearray, not NoneType",
+  "results": {
+    "arc_challenge": {
+      "acc,none": 0.1697952218430034,
+      "acc_norm,none": 0.2226962457337884,
+      "acc_norm_stderr,none": 0.012158314774829888,
+      "acc_stderr,none": 0.010971775157784207,
+      "alias": "arc_challenge",
+      "name": "arc_challenge",
+      "sample_len": 1172
+    },
+    "arc_easy": {
+      "acc,none": 0.3409090909090909,
+      "acc_norm,none": 0.3312289562289562,
+      "acc_norm_stderr,none": 0.009657641311350737,
+      "acc_stderr,none": 0.009726579593423981,
+      "alias": "arc_easy",
+      "name": "arc_easy",
+      "sample_len": 2376
+    },
+    "hellaswag": {
+      "acc,none": 0.2669786895040829,
+      "acc_norm,none": 0.2756423023302131,
+      "acc_norm_stderr,none": 0.004459241474518529,
+      "acc_stderr,none": 0.004414770331224373,
+      "alias": "hellaswag",
+      "name": "hellaswag",
+      "sample_len": 10042
+    },
+    "piqa": {
+      "acc,none": 0.5625680087051143,
+      "acc_norm,none": 0.5554951033732318,
+      "acc_norm_stderr,none": 0.011593746871584282,
+      "acc_stderr,none": 0.011574126069682592,
+      "alias": "piqa",
+      "name": "piqa",
+      "sample_len": 1838
+    }
+  },
+  "transformers_version": "5.14.1",
+  "upper_git_hash": null,
+  "versions": {
+    "arc_challenge": 1.0,
+    "arc_easy": 1.0,
+    "hellaswag": 1.0,
+    "piqa": 1.0
+  }
+}

export_manifest.json

@@ -0,0 +1,91 @@
+{
+  "base_token_presentations": 1500000000,
+  "checkpoint": "artifacts/runs/natural20b_pilot_muon_b32/checkpoint-tokens-0500000000",
+  "checkpoint_model_sha256": "c795e5b65cee129c5c1261cb12108d7ff814203342595b86409d1a56ae08ef7b",
+  "created_at_utc": "2026-08-14T17:23:09.130863+00:00",
+  "cumulative_token_presentations": 2000170752,
+  "files": {
+    "LICENSE": {
+      "bytes": 11357,
+      "sha256": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4"
+    },
+    "NOTICE": {
+      "bytes": 809,
+      "sha256": "2528432a74beebc46af7f34753b2d280d7ac09d7b4caf5ea9e9b4b0710f0912d"
+    },
+    "README.md": {
+      "bytes": 4190,
+      "sha256": "af0b0c22af1207faa91fb5209a9cf38d496c174f146abb18d27558d1e8a1c036"
+    },
+    "TRAINING_DATA.md": {
+      "bytes": 2693,
+      "sha256": "1d3986e1cea683d91e3e16385a4c184f1671f11f0ae73b9fb32c5da442f200ef"
+    },
+    "chat_template.jinja": {
+      "bytes": 184,
+      "sha256": "df23563531f85f90a16699e28d1fe2458f8d6e46e9a4ab61a9ce0d3b1459bf8c"
+    },
+    "config.json": {
+      "bytes": 1877,
+      "sha256": "96d46e3e2ac6df31c52613f981ade8bded88ba11f19ff6269ade373d04772a6b"
+    },
+    "configuration_microloop.py": {
+      "bytes": 8042,
+      "sha256": "8c2777d90977ca1c7051b68ca4abc7da8bc978ffe9c64c34fe66a4d5e3febd2c"
+    },
+    "evaluation/arithmark_3.json": {
+      "bytes": 4693,
+      "sha256": "6c9970b0eb850a4bd8b5ae0ed377af7d829ec8f40316f92bdc79f6e2e64d17a9"
+    },
+    "evaluation/open_slm.json": {
+      "bytes": 9549,
+      "sha256": "4794f515a80ace8134870e8ac4ca89ad62818d7a53d639473de5189df2786d74"
+    },
+    "generation_config.json": {
+      "bytes": 250,
+      "sha256": "c086c408dba55e547bac299549a8b7713c768cb94a52fd2043429b1564d76094"
+    },
+    "model.safetensors": {
+      "bytes": 39041336,
+      "sha256": "c795e5b65cee129c5c1261cb12108d7ff814203342595b86409d1a56ae08ef7b"
+    },
+    "modeling_microloop.py": {
+      "bytes": 25617,
+      "sha256": "3dbb5823a62d7f0228a3770411c41503ac3e28ee99bdea5da8ea45fca08acfb7"
+    },
+    "tokenizer.json": {
+      "bytes": 265304,
+      "sha256": "c063fd42efccd8416d0375230767949356088cc430c8e74d4547438479889077"
+    },
+    "tokenizer_config.json": {
+      "bytes": 482,
+      "sha256": "87e2a284910c5738892c66e027fa8d9a9d2a4cdc7311a4e9503cffbb7997eeb8"
+    },
+    "training/corpus_contract.yaml": {
+      "bytes": 4528,
+      "sha256": "08c628840f3bac74fc400a4a70a6e38057880d68dc5914d03a9e2803130b946a"
+    },
+    "training/packed_metadata.json": {
+      "bytes": 1434,
+      "sha256": "5bc3d4e571fd223eef0523f210bbf17d4c5000f1366b84da11d014ce47228a68"
+    },
+    "training/provenance.json": {
+      "bytes": 2250,
+      "sha256": "43156355a8da05338b86ff91736d74efe5cb58a8d194a9214ab14d809f6d9592"
+    }
+  },
+  "format_version": 1,
+  "license": "apache-2.0",
+  "model_id_placeholder": "juinron/Hummingbird-V1",
+  "parameter_count": 9592720,
+  "scores": {
+    "arc_challenge": 22.26962457337884,
+    "arc_easy": 33.12289562289562,
+    "arithmark_3": 36.1,
+    "hellaswag": 27.56423023302131,
+    "int_index": 7.5980843240706175,
+    "piqa": 55.549510337323184
+  },
+  "tokenizer": "artifacts/runs/e3_tokenizers/tok_4k_digit",
+  "tokens_seen": 500170752
+}

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "do_sample": false,
+  "eos_token_id": 2,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 0,
+  "transformers_version": "5.14.1",
+  "use_cache": false
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c795e5b65cee129c5c1261cb12108d7ff814203342595b86409d1a56ae08ef7b
+size 39041336

modeling_microloop.py

@@ -0,0 +1,553 @@
+"""Small, HF-compatible causal bootstrap model used by the M0 gate.
+
+The diffusion objective and sampler are deliberately separate modules.  This model
+provides the shared transformer backbone and a causal forward path so that the
+project can validate shape correctness, parameter accounting, and reproducibility
+before any expensive data work begins.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from torch import Tensor, nn
+from torch.nn import functional as F
+
+from .configuration_microloop import MicroLoopConfig
+
+try:
+    from transformers import PreTrainedModel
+    from transformers.generation import GenerationMixin
+    from transformers.utils import ModelOutput
+except ImportError:  # pragma: no cover - only used in a minimal environment.
+
+    class GenerationMixin:  # type: ignore[no-redef]
+        pass
+
+    class ModelOutput:  # type: ignore[no-redef]
+        pass
+
+    class PreTrainedModel(nn.Module):  # type: ignore[no-redef]
+        config_class = MicroLoopConfig
+        base_model_prefix = "microloop"
+
+        def __init__(self, config: MicroLoopConfig) -> None:
+            super().__init__()
+            self.config = config
+
+
+@dataclass
+class MicroLoopCausalLMOutput(ModelOutput):
+    """Minimal output object with both attribute and mapping-style access."""
+
+    logits: Tensor
+    loss: Optional[Tensor] = None
+    hidden_states: Optional[Tensor] = None
+    loop_applications: Optional[int] = None
+
+    def __getitem__(self, key: str):
+        return getattr(self, key)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+
+    def forward(self, hidden_states: Tensor) -> Tensor:
+        # Explicit computation rather than the fused ``F.rms_norm`` kernel:
+        # the fused kernel selects implementations based on process-level state
+        # and produces context-dependent numerics inside the training process
+        # (the 2026-08-05 provenance incident).  This explicit path is
+        # deterministic everywhere; the small speed cost is acceptable here.
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        return (hidden_states / torch.sqrt(variance + self.eps)) * self.weight
+
+
+def _rotate_half(x: Tensor) -> Tensor:
+    x_even = x[..., ::2]
+    x_odd = x[..., 1::2]
+    return torch.stack((-x_odd, x_even), dim=-1).flatten(-2)
+
+
+def _rope_tables(max_position: int, head_dim: int, theta: float) -> tuple[Tensor, Tensor]:
+    """Build interleaved rotary tables once, in float32 for stable reuse."""
+
+    inv_freq = 1.0 / (theta ** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim))
+    positions = torch.arange(max_position, device=inv_freq.device, dtype=torch.float32)
+    angles = positions.unsqueeze(-1) * inv_freq
+    angles = torch.stack((angles, angles), dim=-1).flatten(-2)
+    return angles.cos(), angles.sin()
+
+
+def _apply_rope(q: Tensor, k: Tensor, cos: Tensor, sin: Tensor) -> tuple[Tensor, Tensor]:
+    """Apply cached interleaved rotary embeddings to query and key tensors."""
+
+    return q * cos + _rotate_half(q) * sin, k * cos + _rotate_half(k) * sin
+
+
+class GroupedQueryAttention(nn.Module):
+    """Grouped-query self-attention with explicit Q/K/V projections.
+
+    Attention uses an explicit scaled-dot-product implementation (matmul +
+    softmax) rather than ``F.scaled_dot_product_attention`` because the fused
+    kernels select implementations based on process-level state and produce
+    deterministic but context-dependent results: evaluations inside the
+    training process then disagree with evaluations of the same saved
+    checkpoint in a fresh process (the 2026-08-05 provenance incident).  The
+    explicit math path is deterministic everywhere at the cost of a small
+    amount of speed, which is acceptable at this model size.
+    """
+
+    def __init__(self, config: MicroLoopConfig) -> None:
+        super().__init__()
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.head_dim = config.head_dimension
+        self.num_groups = self.num_heads // self.num_key_value_heads
+        self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(
+            config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        self.output_gate = (
+            nn.Linear(config.hidden_size, self.num_heads, bias=False)
+            if config.attention_output_gate
+            else None
+        )
+        self.rope_theta = config.rope_theta
+        self.attention_implementation = config.attention_implementation
+        if config.qk_norm == "per_head":
+            # One affine scale is shared by all Q heads and one by all K heads;
+            # normalization itself is applied independently over each head.
+            self.q_norm: RMSNorm | None = RMSNorm(self.head_dim, config.rms_norm_eps)
+            self.k_norm: RMSNorm | None = RMSNorm(self.head_dim, config.rms_norm_eps)
+        else:
+            self.q_norm = None
+            self.k_norm = None
+
+    def forward(
+        self,
+        hidden_states: Tensor,
+        attention_mask: Tensor | None = None,
+        position_ids: Tensor | None = None,
+        rope_embeddings: tuple[Tensor, Tensor] | None = None,
+    ) -> Tensor:
+        batch, sequence, _ = hidden_states.shape
+        q = self.q_proj(hidden_states).view(batch, sequence, self.num_heads, self.head_dim)
+        k = self.k_proj(hidden_states).view(
+            batch, sequence, self.num_key_value_heads, self.head_dim
+        )
+        v = self.v_proj(hidden_states).view(
+            batch, sequence, self.num_key_value_heads, self.head_dim
+        )
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+        if position_ids is None:
+            position_ids = torch.arange(sequence, device=hidden_states.device).expand(batch, -1)
+        if position_ids.shape != (batch, sequence):
+            raise ValueError(
+                f"position_ids must have shape [batch, sequence], got {tuple(position_ids.shape)}"
+            )
+        if rope_embeddings is None:
+            table_cos, table_sin = _rope_tables(sequence, self.head_dim, self.rope_theta)
+            cos = table_cos[position_ids].unsqueeze(1).to(q.dtype)
+            sin = table_sin[position_ids].unsqueeze(1).to(q.dtype)
+        else:
+            cos, sin = rope_embeddings
+        q, k = _apply_rope(q, k, cos, sin)
+        if self.q_norm is not None:
+            q = self.q_norm(q)
+            assert self.k_norm is not None
+            k = self.k_norm(k)
+
+        if self.num_heads % self.num_key_value_heads != 0:
+            raise ValueError(
+                f"num_heads {self.num_heads} must divide num_key_value_heads "
+                f"{self.num_key_value_heads}"
+            )
+        if attention_mask is None:
+            visible: Tensor | None = None
+            is_causal = True
+        else:
+            # A 2-D mask uses the conventional HF meaning: one means visible.
+            if attention_mask.shape == (batch, sequence):
+                causal = torch.tril(
+                    torch.ones(sequence, sequence, device=q.device, dtype=torch.bool)
+                )
+                visible = causal.unsqueeze(0).unsqueeze(0) & attention_mask.bool().unsqueeze(
+                    1
+                ).unsqueeze(2)
+                is_causal = False
+            elif attention_mask.shape == (batch, sequence, sequence):
+                visible = attention_mask.bool().unsqueeze(1)
+                is_causal = False
+            else:
+                raise ValueError(
+                    "attention_mask must have shape [batch, sequence] or "
+                    "[batch, sequence, sequence], "
+                    f"got {tuple(attention_mask.shape)}"
+                )
+        # GQA: repeat the KV heads so every query head has its own K/V.
+        repeat = self.num_heads // self.num_key_value_heads
+        if repeat > 1:
+            k = k.repeat_interleave(repeat, dim=1)
+            v = v.repeat_interleave(repeat, dim=1)
+        if self.attention_implementation == "sdpa":
+            attended = F.scaled_dot_product_attention(
+                q, k, v, attn_mask=visible, dropout_p=0.0, is_causal=is_causal
+            )
+        else:
+            scores = torch.matmul(q, k.transpose(-2, -1)) / float(self.head_dim) ** 0.5
+            if is_causal:
+                seq_ids = torch.arange(sequence, device=scores.device)
+                visible = seq_ids.unsqueeze(0) <= seq_ids.unsqueeze(1)
+                visible = visible.expand(batch, self.num_heads, sequence, sequence)
+            if visible is not None:
+                scores = scores.masked_fill(~visible, float("-inf"))
+            probs = torch.softmax(scores, dim=-1)
+            attended = torch.matmul(probs, v)
+        if self.output_gate is not None:
+            gate = F.silu(self.output_gate(hidden_states)).transpose(1, 2).unsqueeze(-1)
+            attended = attended * gate
+        attended = attended.transpose(1, 2).contiguous().view(batch, sequence, -1)
+        return self.o_proj(attended)
+
+
+class MicroLoopBlock(nn.Module):
+    """Pre-norm transformer block with SwiGLU feed-forward network."""
+
+    def __init__(self, config: MicroLoopConfig) -> None:
+        super().__init__()
+        self.attn_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.attn = GroupedQueryAttention(config)
+        self.ffn_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.ffn_gate = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.ffn_up = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.ffn_down = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+        self.use_attn_residuals = config.attn_res_block_size is not None
+        if self.use_attn_residuals:
+            self.attn_res_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+            self.ffn_res_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+            self.attn_res_proj = nn.Linear(config.hidden_size, 1, bias=False)
+            self.ffn_res_proj = nn.Linear(config.hidden_size, 1, bias=False)
+        clamp = dict(config.swiglu_clamp)
+        self.swiglu_clamp_enabled = bool(clamp.get("enabled", False))
+        self.swiglu_linear_min = float(clamp.get("linear_min", -10.0))
+        self.swiglu_linear_max = float(clamp.get("linear_max", 10.0))
+        self.swiglu_gate_max = float(clamp.get("gate_max", 10.0))
+
+    def forward(
+        self,
+        hidden_states: Tensor,
+        attention_mask: Tensor | None = None,
+        position_ids: Tensor | None = None,
+        rope_embeddings: tuple[Tensor, Tensor] | None = None,
+        block_residuals: list[Tensor] | None = None,
+    ) -> Tensor:
+        if self.use_attn_residuals and block_residuals:
+            residual_stack = torch.stack(block_residuals, dim=-2)
+            scores = torch.cat(
+                [self.attn_res_proj(self.attn_res_norm(state)) for state in block_residuals], dim=-1
+            )
+            hidden_states = hidden_states + (
+                torch.softmax(scores, dim=-1).unsqueeze(-1) * residual_stack
+            ).sum(dim=-2)
+        hidden_states = hidden_states + self.attn(
+            self.attn_norm(hidden_states),
+            attention_mask,
+            position_ids,
+            rope_embeddings,
+        )
+        if self.use_attn_residuals and block_residuals:
+            residual_stack = torch.stack(block_residuals, dim=-2)
+            scores = torch.cat(
+                [self.ffn_res_proj(self.ffn_res_norm(state)) for state in block_residuals], dim=-1
+            )
+            hidden_states = hidden_states + (
+                torch.softmax(scores, dim=-1).unsqueeze(-1) * residual_stack
+            ).sum(dim=-2)
+        ffn_input = self.ffn_norm(hidden_states)
+        gate_linear = self.ffn_gate(ffn_input)
+        up_linear = self.ffn_up(ffn_input)
+        if self.swiglu_clamp_enabled:
+            gate_linear = gate_linear.clamp(self.swiglu_linear_min, self.swiglu_linear_max)
+            up_linear = up_linear.clamp(self.swiglu_linear_min, self.swiglu_linear_max)
+        gate = F.silu(gate_linear)
+        if self.swiglu_clamp_enabled:
+            gate = gate.clamp(max=self.swiglu_gate_max)
+        ffn_output = self.ffn_down(gate * up_linear)
+        return hidden_states + ffn_output
+
+
+class MicroLoopPreTrainedModel(PreTrainedModel):
+    config_class = MicroLoopConfig
+    base_model_prefix = "microloop"
+
+
+class MicroLoopForDiffusionLM(MicroLoopPreTrainedModel, GenerationMixin):
+    """Backbone plus tied output head for causal bootstrap and diffusion training."""
+
+    _tied_weights_keys = {"lm_head.weight": "embed_tokens.weight"}
+
+    def __init__(self, config: MicroLoopConfig) -> None:
+        config.validate()
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList(
+            [MicroLoopBlock(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.mtp_proj = (
+            nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+            if config.mtp_enabled
+            else None
+        )
+        rope_cos, rope_sin = _rope_tables(
+            config.max_position_embeddings, config.head_dimension, config.rope_theta
+        )
+        # PERSISTENT buffers on purpose: transformers 5.x ``from_pretrained``
+        # re-initializes non-persistent buffers that are missing from the
+        # checkpoint (``_initialize_missing_keys`` -> ``initialize_weights``),
+        # overwriting the rotary tables with garbage.  That made every fresh-
+        # process evaluation of saved checkpoints compute with corrupted rope
+        # tables (the historical "external collapse" at ~0.03 accuracy was this
+        # artifact).  Persisting the tables makes the saved checkpoint carry
+        # exactly the tables used in training.
+        self.register_buffer("rope_cos", rope_cos, persistent=True)
+        self.register_buffer("rope_sin", rope_sin, persistent=True)
+        if config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+        # Canonical HF pattern: post_init() installs all_tied_weights_keys and
+        # dispatches _initialize_weights per module.
+        self.post_init()
+
+    def _reset_rope_buffers(self) -> None:
+        """Recompute the rotary tables from the config (they are deterministic)."""
+
+        rope_cos, rope_sin = _rope_tables(
+            self.config.max_position_embeddings,
+            self.config.head_dimension,
+            self.config.rope_theta,
+        )
+        self.rope_cos.copy_(rope_cos)
+        self.rope_sin.copy_(rope_sin)
+
+    def _initialize_weights(self, module: nn.Module, is_custom_code: bool = False) -> None:
+        # The caller controls the RNG through seed_everything; this method performs
+        # no hidden reseeding and is therefore reproducible by construction.
+        if getattr(module, "_is_hf_initialized", False):
+            return
+        if module is self:
+            # The main module owns the rotary tables.  transformers 5.x
+            # re-initializes buffers that are missing from a loaded checkpoint
+            # ("_initialize_missing_keys"), which zeroes/garbles the tables;
+            # restore the deterministic canonical tables here instead.
+            self._reset_rope_buffers()
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if getattr(module, "bias", None) is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, RMSNorm):
+            nn.init.ones_(module.weight)
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embed_tokens
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.lm_head
+
+    def set_input_embeddings(self, value: nn.Embedding) -> None:
+        self.embed_tokens = value
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+
+    def forward(
+        self,
+        input_ids: Tensor | None = None,
+        inputs_embeds: Tensor | None = None,
+        attention_mask: Tensor | None = None,
+        document_ids: Tensor | None = None,
+        position_ids: Tensor | None = None,
+        labels: Tensor | None = None,
+        logit_mask: Tensor | None = None,
+        mtp_loss_weight: float = 0.0,
+        loop_count: int = 1,
+        output_hidden_states: bool = False,
+        **_: object,
+    ) -> MicroLoopCausalLMOutput:
+        if (input_ids is None) == (inputs_embeds is None):
+            raise ValueError("exactly one of input_ids or inputs_embeds must be provided")
+        if loop_count < 1:
+            raise ValueError("loop_count must be at least one")
+        if input_ids is not None:
+            if input_ids.dim() != 2:
+                raise ValueError(
+                    f"input_ids must have shape [batch, sequence], got {input_ids.dim()}-D"
+                )
+            batch, sequence = input_ids.shape
+        else:
+            if inputs_embeds is None or inputs_embeds.dim() != 3:
+                raise ValueError(
+                    "inputs_embeds must have shape [batch, sequence, hidden], got "
+                    f"{None if inputs_embeds is None else inputs_embeds.dim()}-D"
+                )
+            batch, sequence, _ = inputs_embeds.shape
+        device = (input_ids if input_ids is not None else inputs_embeds).device
+        if document_ids is not None:
+            if document_ids.shape != (batch, sequence):
+                raise ValueError("document_ids must have the same shape as the input")
+            if attention_mask is not None and attention_mask.dim() != 2:
+                raise ValueError(
+                    "document_ids cannot be combined with a precomputed attention mask"
+                )
+            causal = torch.tril(torch.ones(sequence, sequence, dtype=torch.bool, device=device))
+            valid = (
+                attention_mask.bool()
+                if attention_mask is not None
+                else torch.ones((batch, sequence), dtype=torch.bool, device=device)
+            )
+            attention_mask = (
+                causal.unsqueeze(0)
+                & document_ids.unsqueeze(2).eq(document_ids.unsqueeze(1))
+                & valid.unsqueeze(1)
+                & valid.unsqueeze(2)
+            )
+        if position_ids is None:
+            position_ids = torch.arange(sequence, device=device).expand(batch, -1)
+        rope_dtype = (
+            torch.get_autocast_dtype("cuda")
+            if device.type == "cuda" and torch.is_autocast_enabled("cuda")
+            else self.embed_tokens.weight.dtype
+        )
+        rope_embeddings = (
+            self.rope_cos[position_ids].unsqueeze(1).to(rope_dtype),
+            self.rope_sin[position_ids].unsqueeze(1).to(rope_dtype),
+        )
+        hidden_states = self.embed_tokens(input_ids) if input_ids is not None else inputs_embeds
+        loop_layers = set(self.config.looping.get("layers", [4, 5, 6]))
+        block_residuals: list[Tensor] = []
+        block_size = self.config.attn_res_block_size
+        total_applications = 0
+        for layer_number, layer in enumerate(self.layers, start=1):
+            if block_size is not None and (layer_number - 1) % block_size == 0:
+                block_residuals.append(hidden_states)
+            prior_block_residuals = block_residuals[:-1] if block_size is not None else None
+            repetitions = loop_count if layer_number in loop_layers else 1
+            for _ in range(repetitions):
+                hidden_states = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    rope_embeddings=rope_embeddings,
+                    block_residuals=prior_block_residuals,
+                )
+                total_applications += 1
+        hidden_states = self.norm(hidden_states)
+        if logit_mask is not None:
+            # Diffusion training: only the masked positions carry loss, so the
+            # output head runs on those rows instead of the full sequence.
+            if labels is not None:
+                raise ValueError("logit_mask cannot be combined with labels")
+            if logit_mask.shape != (batch, sequence):
+                raise ValueError("logit_mask must have the same shape as the input")
+            flat = hidden_states.reshape(-1, self.config.hidden_size)
+            logits = self.lm_head(flat[logit_mask.reshape(-1)])
+        else:
+            logits = self.lm_head(hidden_states)
+        loss = None
+        if labels is not None:
+            if labels.shape != (batch, sequence):
+                raise ValueError("labels must have the same shape as the input")
+            if labels.size(1) < 2:
+                raise ValueError("causal training requires sequences with at least two tokens")
+            # Causal next-token prediction: position t predicts the label at t + 1.
+            loss = F.cross_entropy(
+                logits[:, :-1, :].reshape(-1, logits.size(-1)),
+                labels[:, 1:].reshape(-1),
+                ignore_index=-100,
+            )
+            if mtp_loss_weight:
+                if self.mtp_proj is None:
+                    raise ValueError("mtp_loss_weight requires mtp_enabled=true")
+                if mtp_loss_weight < 0:
+                    raise ValueError("mtp_loss_weight must be non-negative")
+                mtp_logits = self.lm_head(self.mtp_proj(hidden_states[:, :-2, :]))
+                mtp_loss = F.cross_entropy(
+                    mtp_logits.reshape(-1, mtp_logits.size(-1)),
+                    labels[:, 2:].reshape(-1),
+                    ignore_index=-100,
+                )
+                loss = loss + float(mtp_loss_weight) * mtp_loss
+        return MicroLoopCausalLMOutput(
+            logits=logits,
+            loss=loss,
+            hidden_states=hidden_states if output_hidden_states else None,
+            loop_applications=total_applications,
+        )
+
+    @torch.no_grad()
+    def generate_greedy(
+        self, input_ids: Tensor, max_new_tokens: int, eos_token_id: int | None = None
+    ) -> Tensor:
+        """Small causal smoke decoder; diffusion sampling belongs in ``sampler.py``."""
+
+        return self.generate_causal(
+            input_ids, max_new_tokens=max_new_tokens, eos_token_id=eos_token_id, do_sample=False
+        )
+
+    @torch.no_grad()
+    def generate_causal(
+        self,
+        input_ids: Tensor,
+        *,
+        max_new_tokens: int,
+        eos_token_id: int | None = None,
+        do_sample: bool = False,
+        temperature: float = 1.0,
+        top_k: int | None = None,
+    ) -> Tensor:
+        """Generate a batched causal continuation for M2 validation and serving.
+
+        This deliberately recomputes the context on every step. KV caching is a
+        later optimization; keeping this reference path simple makes M2 output
+        semantics straightforward to test.
+        """
+
+        if max_new_tokens < 0:
+            raise ValueError("max_new_tokens must be non-negative")
+        if temperature <= 0:
+            raise ValueError("temperature must be positive")
+        if top_k is not None and top_k <= 0:
+            raise ValueError("top_k must be positive when supplied")
+        generated = input_ids
+        finished = torch.zeros(input_ids.size(0), dtype=torch.bool, device=input_ids.device)
+        for _ in range(max_new_tokens):
+            next_logits = self(generated).logits[:, -1, :]
+            if do_sample:
+                next_logits = next_logits / temperature
+                if top_k is not None and top_k < next_logits.size(-1):
+                    threshold = torch.topk(next_logits, top_k, dim=-1).values[:, -1:]
+                    next_logits = next_logits.masked_fill(next_logits < threshold, float("-inf"))
+                next_token = torch.multinomial(torch.softmax(next_logits, dim=-1), 1)
+            else:
+                next_token = next_logits.argmax(dim=-1, keepdim=True)
+            if eos_token_id is not None:
+                next_token = torch.where(
+                    finished.unsqueeze(1),
+                    torch.full_like(next_token, eos_token_id),
+                    next_token,
+                )
+                finished |= next_token.squeeze(1).eq(eos_token_id)
+            generated = torch.cat((generated, next_token), dim=1)
+            if eos_token_id is not None and bool(finished.all()):
+                break
+        return generated

tokenizer_config.json

@@ -0,0 +1,21 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<bos>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<eos>",
+  "extra_special_tokens": [
+    "<system>",
+    "<user>",
+    "<assistant>",
+    "<turn_end>",
+    "<doc_end>"
+  ],
+  "is_local": true,
+  "local_files_only": false,
+  "mask_token": "<mask>",
+  "model_max_length": 2048,
+  "pad_token": "<pad>",
+  "padding_side": "right",
+  "tokenizer_class": "TokenizersBackend",
+  "truncation_side": "right"
+}

training/corpus_contract.yaml

@@ -0,0 +1,142 @@
+# Frost all-natural 20B-token pretraining corpus contract (pinned 2026-08-14).
+#
+# This contract is intentionally separate from final2b_corpus_contract.yaml: no
+# generated arithmetic, rationale, tutorial, or benchmark-derived text is
+# admitted. Public Open SLM tasks remain audit-only and are protected before
+# global exact/near deduplication.
+version: 1
+project: frost-10m
+purpose: natural_20b_unique_token_pretraining_corpus
+frozen: 2026-08-14
+seed: 42
+target:
+  unique_train_tokens: 20000000000
+  train_presentations: 20000000000
+  gross_selected_tokens: 22000000000
+  split_policy: document_sha256_mod_100
+  validation_buckets: [0]
+  held_out_buckets: [1]
+  train_buckets: 2-99
+
+# Shares are the desired packed-train mix. Gross quotas carry a 10% reserve for
+# the two held-out hash buckets plus quality and cross-source dedup attrition.
+source_mix:
+  - name: educational_web
+    share: 0.50
+    tokens_quota: 11000000000
+    fill: [natural_fineweb_edu]
+  - name: broad_web
+    share: 0.20
+    tokens_quota: 4400000000
+    fill: [natural_dclm]
+  - name: high_quality_web
+    share: 0.10
+    tokens_quota: 2200000000
+    fill: [natural_fineweb_hq]
+  - name: textbook_exposition
+    share: 0.15
+    tokens_quota: 3300000000
+    fill: [natural_cosmopedia_v2]
+  - name: mathematics
+    share: 0.05
+    tokens_quota: 1100000000
+    fill: [natural_finemath_4plus]
+
+# Keep the most curated sources first so keep-first cross-source dedup retains
+# the higher-quality copy. Every revision is an immutable Hub commit observed
+# through the Hugging Face dataset API on 2026-08-14.
+build_order:
+  - natural_finemath_4plus
+  - natural_cosmopedia_v2
+  - natural_fineweb_hq
+  - natural_fineweb_edu
+  - natural_dclm
+
+sources:
+  - name: natural_fineweb_edu
+    row: educational_web
+    status: pinned
+    repo_id: HuggingFaceFW/fineweb-edu
+    revision: 87f09149ef4734204d70ed1d046ddc9ca3f2b8f9
+    config: sample-100BT
+    split: train
+    text_field: text
+    license: odc-by-1.0
+    shuffle_buffer: 10000
+    quality_filter: m9_admission
+  - name: natural_dclm
+    row: broad_web
+    status: pinned
+    repo_id: mlfoundations/dclm-baseline-1.0
+    revision: a3b142c183aebe5af344955ae20836eb34dcf69b
+    config: default
+    split: train
+    text_field: text
+    license: cc-by-4.0
+    shuffle_buffer: 10000
+    quality_filter: m9_admission
+  - name: natural_fineweb_hq
+    row: high_quality_web
+    status: pinned
+    repo_id: epfml/FineWeb-HQ
+    revision: e58199cdd52438d94405df1a4d8630cc5f13bf84
+    config: default
+    split: train
+    text_field: text
+    license: odc-by-1.0
+    shuffle_buffer: 10000
+    quality_filter: m9_admission
+  - name: natural_cosmopedia_v2
+    row: textbook_exposition
+    status: pinned
+    repo_id: HuggingFaceTB/smollm-corpus
+    revision: 3ba9d605774198c5868892d7a8deda78031a781f
+    config: cosmopedia-v2
+    split: train
+    text_field: text
+    license: odc-by-1.0
+    shuffle_buffer: 10000
+    quality_filter: m9_admission
+  - name: natural_finemath_4plus
+    row: mathematics
+    status: pinned
+    repo_id: HuggingFaceTB/finemath
+    revision: e92b25a616738fe95dc186b64dfb19f9c8525594
+    config: finemath-4plus
+    split: train
+    text_field: text
+    license: odc-by-1.0
+    shuffle_buffer: 10000
+    quality_filter: m9_admission
+
+quality:
+  min_chars: 40
+  max_chars: 100000
+  max_symbol_ratio: 0.35
+  min_ascii_alpha_ratio: 0.45
+  reject_secrets: true
+deduplication:
+  exact: document_sha256_canonical
+  near_duplicate:
+    method: 32-permutation MinHash / 8-band LSH
+    shingle_width: 5
+    threshold: 0.80
+    confirmation: exact Jaccard
+  cross_source: global_keep_first_in_build_order
+  cross_split: keep_first_over_union_before_hash_split
+protection_index:
+  public_suite: artifacts/m9_protection/protection.manifest.json
+  internal_dev: artifacts/m9_internal_dev
+  terse_choice: artifacts/m10_terse/terse_dev.jsonl
+pilot:
+  target_train_tokens: 1000000000
+  purpose: test mix, stability, loss slope, and internal-dev transfer before 20B
+  checkpoint_tokens: [250000000, 500000000, 750000000, 1000000000]
+scaling_gate:
+  local_in_ram_minhash_max_tokens: 2000000000
+  full_20b_requirement: distributed or disk-sharded MinHash before packing
+  reason: avoid silently exhausting workstation RAM with a tens-of-millions-document index
+selection_rule: >-
+  Choose LR and continuation only on training stability, held-out loss,
+  mathematics-dev loss, terse-choice accuracy, and non-public generation probes.
+  Public Open SLM results are reported only after the pilot decision is frozen.

training/packed_metadata.json

@@ -0,0 +1,50 @@
+{
+  "document_end_id": 8,
+  "elapsed_seconds": 541.1,
+  "format_version": 3,
+  "held_out_percent": 1,
+  "mix_policy": "token_balanced",
+  "per_source_tokens_all_splits": {
+    "natural_cosmopedia_v2": 168663969,
+    "natural_dclm": 211970663,
+    "natural_finemath_4plus": 53640449,
+    "natural_fineweb_edu": 570339212,
+    "natural_fineweb_hq": 109226534
+  },
+  "row_targets": {
+    "broad_web": 0.2,
+    "educational_web": 0.5,
+    "high_quality_web": 0.1,
+    "mathematics": 0.05,
+    "textbook_exposition": 0.15
+  },
+  "row_tokens_all_splits": {
+    "broad_web": 211970663,
+    "educational_web": 570339212,
+    "high_quality_web": 109226534,
+    "mathematics": 53640449,
+    "textbook_exposition": 168663969
+  },
+  "splits": {
+    "held_out": {
+      "document_ends_file": "held_out_document_ends.npy",
+      "documents": 8586,
+      "token_file": "held_out_tokens.int32",
+      "tokens": 11432331
+    },
+    "train": {
+      "document_ends_file": "train_document_ends.npy",
+      "documents": 824997,
+      "token_file": "train_tokens.int32",
+      "tokens": 1091660174
+    },
+    "validation": {
+      "document_ends_file": "validation_document_ends.npy",
+      "documents": 8437,
+      "token_file": "validation_tokens.int32",
+      "tokens": 10748322
+    }
+  },
+  "tokenizer_path": "D:\\llm\\frost\\artifacts\\runs\\e3_tokenizers\\tok_4k_digit",
+  "validation_percent": 1
+}

training/provenance.json

@@ -0,0 +1,65 @@
+{
+  "base_checkpoint": "artifacts/runs/variant_v1_muon_500m/checkpoint-tokens-1500000000",
+  "base_token_presentations": 1500000000,
+  "checkpoint_selection": {
+    "evaluated_continuation_checkpoints": [
+      250000000,
+      500000000,
+      750000000,
+      1000000000
+    ],
+    "reason": "highest observed chance-normalized Open SLM Intelligence Index",
+    "selected_continuation_checkpoint": 500000000,
+    "used_public_evaluations": true
+  },
+  "continuation_checkpoint_metrics": {
+    "elapsed_seconds": 4661.750659100002,
+    "gradient_norm": 0.7176291942596436,
+    "learning_rate": 0.0003,
+    "step": 1908,
+    "tokens_seen": 500170752,
+    "train_loss": 2.8427783250808716,
+    "validation_loss": 2.7994241237640383
+  },
+  "continuation_token_presentations": 500170752,
+  "cumulative_token_presentations": 2000170752,
+  "format_version": 1,
+  "release_name": "Hummingbird-V1",
+  "source_checkpoint": "artifacts/runs/natural20b_pilot_muon_b32/checkpoint-tokens-0500000000",
+  "training_config": {
+    "adam_beta1": 0.9,
+    "adam_beta2": 0.95,
+    "adam_eps": 1e-08,
+    "compile": false,
+    "context_length": 512,
+    "corpus_path": "artifacts\\natural20b_pilot\\packed",
+    "deterministic": false,
+    "device": null,
+    "gradient_accumulation_steps": 16,
+    "init_from": "artifacts/runs/variant_v1_muon_500m/checkpoint-tokens-1500000000",
+    "learning_rate": 0.0003,
+    "lr_decay_fraction": 0.15,
+    "lr_min_ratio": 0.1,
+    "lr_schedule": "wsd",
+    "lr_warmup_steps": 100,
+    "max_grad_norm": 1.0,
+    "micro_batch_size": 32,
+    "model_config_path": "configs\\model_4k_14l_sdpa_qk_clamp.yaml",
+    "mtp_final_loss_weight": null,
+    "mtp_loss_weight": 0.0,
+    "num_workers": 0,
+    "optimizer": "muon",
+    "pin_memory": true,
+    "precision": "bf16",
+    "prepared_data_dir": "artifacts\\natural20b_pilot\\packed",
+    "run_dir": "artifacts\\runs\\natural20b_pilot_muon_b32",
+    "save_every_steps": 500,
+    "save_every_tokens": 250000000,
+    "seed": 42,
+    "tokenizer_path": "artifacts\\runs\\e3_tokenizers\\tok_4k_digit",
+    "tokens_target": 1000000000,
+    "validate_every_steps": 100,
+    "validation_tokens": 65536,
+    "weight_decay": 0.1
+  }
+}