Release Fields 300M PG-19 checkpoint v0.1.0

ENVIRONMENT.txt

@@ -0,0 +1,18 @@
+generated_utc=2026-06-29T22:53:25Z
+kernel=Linux 6.8.0-1046-nvidia x86_64 GNU/Linux
+
+NVIDIA H100 80GB HBM3, 580.105.08, 81559 MiB
+
+python 3.10.12 (main, Jan 26 2026, 14:55:28) [GCC 11.4.0]
+platform Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35
+torch 2.7.0
+triton 3.3.0
+mamba_ssm 2.3.2.post1
+causal_conv1d 1.6.2.post1
+transformers 5.12.1
+tokenizers 0.22.2
+huggingface_hub 1.21.0
+safetensors 0.8.0
+torch_cuda 12.8
+torch_cuda_available True
+torch_gpu NVIDIA H100 80GB HBM3

EQUIVALENCE_REPORT.json

@@ -0,0 +1,13 @@
+{
+  "abs_loss_diff": 0.0,
+  "artifact": "/home/ubuntu/fields_hf_export",
+  "backend_name": "final_field_pcaf_on",
+  "canonical_sha256": "0454647a7f667cdf73f45029eee0f525c58c0788ecd433ec1a4db8c3937ac848",
+  "checkpoint": "/home/ubuntu/pcaf_runs/field_fusion_final_closure_run/quality/seed0_1234/field_official_18f2m4r_pcaf_on/final_bf16.pt",
+  "device": "cuda",
+  "max_abs_logits_diff": 0.0,
+  "pcaf_enabled": true,
+  "sequence_length": 128,
+  "state_dict_exact": true,
+  "status": "PASS"
+}

README.md

@@ -0,0 +1,76 @@
+---
+library_name: fields-lm
+license: apache-2.0
+pipeline_tag: text-generation
+tags:
+- pytorch
+- causal-lm
+- long-context
+- state-space-model
+- mamba
+- research
+language:
+- en
+datasets:
+- emozilla/pg19
+---
+
+# Fields 300M PG-19
+
+This repository contains the first public checkpoint for **Fields LM**, an experimental causal language-model architecture combining native Field recurrence, displaced readout, PCAF successor memory, local refresh attention, and Mamba-2 editors.
+
+## Model details
+
+- architecture: Fields 18F/2M/4R + PCAF;
+- scale: approximately 300M parameters;
+- vocabulary: 16,384-token BPE;
+- training data: controlled sample of PG-19;
+- training budget: 49,152,000 tokens;
+- training context: 2,048 tokens;
+- evaluated contexts: 2K, 8K, 16K, 32K, and 64K;
+- weights: `safetensors`;
+- runtime: install from the official GitHub repository.
+
+This is **not** a frontier-scale assistant and is not instruction-tuned. Long-context values beyond 2K measure evaluation-time extrapolation, not training at those lengths.
+
+## Installation
+
+```bash
+pip install "git+https://github.com/Multisymboliccore/fields-lm.git"
+```
+
+Install the validated CUDA dependencies for Mamba-2 and causal convolution as documented in the GitHub repository.
+
+## Loading
+
+```python
+import torch
+from fields_official import FieldsHubModel
+
+model = FieldsHubModel.from_pretrained(
+    "Multisymboliccore/fields-300m-pg19",
+    map_location="cpu",
+)
+model = model.to(device="cuda", dtype=torch.bfloat16).eval()
+```
+
+## Intended use
+
+The checkpoint is intended for architecture research, reproducibility, controlled ablations, long-context analysis, and further pretraining or fine-tuning by qualified users.
+
+## Limitations
+
+- trained with a comparatively small token budget;
+- English literary-book domain bias from PG-19;
+- not instruction-tuned or safety-tuned;
+- may generate inaccurate, biased, or harmful text;
+- custom CUDA dependencies are required for the validated high-performance path;
+- 64K evaluation does not imply equal quality across all long-context tasks.
+
+## Evaluation
+
+Final three-seed metrics and paper links will be synchronized from the validated release artifacts. Do not infer universal model quality from the architecture benchmark alone.
+
+## License
+
+Apache License 2.0. Dependency licenses continue to apply.

SHA256SUMS.txt

@@ -0,0 +1,9 @@
+6a79821480e8c73bb79d9bf1a0b81247df4dc6269e520de92f1e8768314962a8  ENVIRONMENT.txt
+5c0743bfe273f83b38154b65e86100bdbb68c5c53846443dfa9b5de504aebc03  EQUIVALENCE_REPORT.json
+9b552d68a5346641d7a09f5ee52a1ff857b5567c7be9ff9d57bce5c207bfce3f  README.md
+46c8141eb4fc663a47623e1fe20201092088a31d036ae404d9ed150942d688ca  SOURCE_IDENTITY.txt
+4271b423cf20a96afe91110b6262080aae91eec976fcef9291e33056001464a1  config.json
+30381e15cc6dd247db9d71a17b63832f6db1387de97ba0efcf168c43797820af  model.safetensors
+c982c025453afc73c93ec1fab4598763bef21cfbf22cd6008d03711d4e6210e1  special_tokens_map.json
+928f6b3b935dcded396b05ae955a344eb1690cca244c988877dc6dd4d38e7470  tokenizer.json
+bd1eaa7215f9b95c8bf3161ce09aea0b28769039d7131fab8d8b17cf35cf06d3  tokenizer_config.json

SOURCE_IDENTITY.txt

@@ -0,0 +1,21 @@
+{
+  "canonical_sha256": "0454647a7f667cdf73f45029eee0f525c58c0788ecd433ec1a4db8c3937ac848",
+  "checkpoint": "/home/ubuntu/pcaf_runs/field_fusion_final_closure_run/quality/seed0_1234/field_official_18f2m4r_pcaf_on/final_bf16.pt",
+  "checkpoint_sha256": "e81f2c1338db6a84231818877067a5743a137990e408ad78396fd5f10178f857",
+  "metadata": {
+    "batch": 4,
+    "canonical_sha256": "0454647a7f667cdf73f45029eee0f525c58c0788ecd433ec1a4db8c3937ac848",
+    "data_manifest_sha256": "2d158e9614bf222db8ae8b9a7d4ecd0b5170dbfaf233d4bedb323ee409cc8716",
+    "data_seed": 5678,
+    "dataset": "emozilla/pg19",
+    "model": "field_official_18f2m4r_pcaf_on",
+    "model_seed": 1234,
+    "parameters": 300343080,
+    "suite_version": 3,
+    "train_seq": 2048,
+    "train_tokens": 49152000
+  },
+  "model": "field_official_18f2m4r_pcaf_on",
+  "tokenizer_sha256": "928f6b3b935dcded396b05ae955a344eb1690cca244c988877dc6dd4d38e7470",
+  "weights_sha256": "30381e15cc6dd247db9d71a17b63832f6db1387de97ba0efcf168c43797820af"
+}

config.json

@@ -0,0 +1,79 @@
+{
+  "amp": "bf16",
+  "architectures": [
+    "FieldsHubModel"
+  ],
+  "backend_name": "final_field_pcaf_on",
+  "canonical_sha256": "0454647a7f667cdf73f45029eee0f525c58c0788ecd433ec1a4db8c3937ac848",
+  "embedding_seed": 314159,
+  "library_name": "fields-lm",
+  "model_max_length": 65536,
+  "model_seed": 1234,
+  "model_type": "fields",
+  "pcaf_enabled": true,
+  "safe_serialization": true,
+  "topology_audit": {
+    "block_count": 24,
+    "classes": [
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FusionRefreshBlockV21",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "Mamba2ResidualBlock",
+      "FusionRefreshBlockV21",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FusionRefreshBlockV21",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "FieldBlock",
+      "Mamba2ResidualBlock",
+      "FusionRefreshBlockV21"
+    ],
+    "mamba_count": 2,
+    "mamba_indices": [
+      10,
+      22
+    ],
+    "native_field_count": 18,
+    "native_field_indices": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      6,
+      7,
+      8,
+      9,
+      12,
+      13,
+      14,
+      15,
+      16,
+      18,
+      19,
+      20,
+      21
+    ],
+    "refresh_count": 4,
+    "refresh_indices": [
+      5,
+      11,
+      17,
+      23
+    ],
+    "rejected_module_names": []
+  },
+  "vocab_size": 16384
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30381e15cc6dd247db9d71a17b63832f6db1387de97ba0efcf168c43797820af
+size 1201405352

special_tokens_map.json

@@ -0,0 +1,3 @@
+{
+  "unk_token": "<unk>"
+}

tokenizer_config.json

@@ -0,0 +1,6 @@
+{
+  "clean_up_tokenization_spaces": false,
+  "model_max_length": 65536,
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "<unk>"
+}