Publish reusable guides, calibration banks and authored evaluation suites

README_agent.md

@@ -0,0 +1,219 @@
+# Character Creation: agent interface and evidence contract
+
+This document is the machine-oriented entry point. [README_human.md](README_human.md)
+explains the motivation and workflow. Source: [GitHub](https://github.com/Atlas3DSS/Character-Creation).
+Public data: `Atlas3D/character-creation-big-five` on Hugging Face, repository type `dataset`.
+
+## Resolve and pin before reading
+
+```python
+from huggingface_hub import HfApi
+repo = "Atlas3D/character-creation-big-five"
+revision = HfApi().dataset_info(repo).sha
+print(revision)  # Record this full SHA with every derived result.
+```
+
+Use that SHA for **every** manifest, batch, asset, and raw-shard download in the analysis.
+Do not mix current `main` files with an earlier snapshot. The dataset changes while
+capture and publication continue. An absent raw entry is missing publication coverage,
+not a zero activation. Failed and incomplete source captures do not become successes
+because another stage finished.
+
+## Public files
+
+| File | Contract |
+| --- | --- |
+| `scenarios.jsonl` | Original synthetic prompt collection; stable `id`, `prompt`, design labels and generation provenance. |
+| `manifest.json` | `character-creation.public-big-five.v1`; exact row count, source SHA-256 and exported JSONL SHA-256. |
+| `research_manifest.json` | `character-creation.public-research.v1`; immutable metadata batches, record IDs, hashes, counts and source progress. |
+| `research_batch_<sha256>.jsonl` | Heterogeneous records; dispatch on `kind`, not guessed columns. |
+| `research_raw_manifest.json` | `character-creation.public-research-raw.v1`; separately verified raw-array publication coverage. |
+| `research_raw_<sha256>.tar` | Flat, immutable shard of original NPZ files; no corpus database. |
+| `research_raw_entries_<sha256>.json` | Raw record-to-entry map, exact bytes, original SHA-256, shapes and dtypes. |
+| `research_assets.json` | `character-creation.public-research-assets.v1`; calibration and authored-suite filenames, hashes, sizes and model binding. |
+| `layer_00.npz` … `layer_63.npz` | Small calibration banks: relative-dose scales, trait directions, selected neuron columns and observed statistics. Not the full model. |
+| `persona_*.json` listed by `research_assets.json` | Authored probe suites, public presets, model metadata and reference recipes. |
+
+Default `datasets.load_dataset(repo, name="default", split="train", revision=revision)`
+loads **only the synthetic scenarios**. `train` is a loader convention; all these rows
+are discovery material. Use the research manifests for mixed research records and arrays.
+
+## Read with verification, without inference
+
+Install `huggingface_hub` and `numpy` in your existing Python environment. From the
+source checkout:
+
+```bash
+python persona/persona_read_research.py --revision FULL_HUB_COMMIT_SHA --limit 3
+python persona/persona_read_research.py --revision FULL_HUB_COMMIT_SHA --kind baseline_answer --limit 3
+```
+
+Programmatic use:
+
+```python
+from persona.persona_read_research import Snapshot
+
+snapshot = Snapshot(revision="FULL_HUB_COMMIT_SHA")
+for row in snapshot.records(kind="prefill_activation"):
+    print(row["id"], row["source"]["id"], row["coverage"])
+    break
+```
+
+Raw access is explicit because one shard can be hundreds of MB:
+
+```python
+arrays = snapshot.arrays(record_id)
+peak = arrays["search_peak"]           # [64, 17408], signed maximum-magnitude value
+where = arrays["search_peak_position"] # [64, 17408], encoded prompt position
+total = arrays["search_sum_abs"]       # [64, 17408], absolute activation sum
+mean_abs = total / len(arrays["search_positions"])
+```
+
+The helper verifies SHA-256, retains a fixed Hub revision, reads tar entries without
+extracting paths, and loads NPZ with `allow_pickle=False`. Download only shards needed
+for the question. Preserve record IDs and exact model/runtime/prepared identities in
+derived tables. Check the raw entry's shapes/dtypes; do not assume every record kind
+contains the same arrays.
+
+## Record kinds and joins
+
+`prefill_activation`:
+
+- Stable record `id`; join to the prompt using `source.id`, not row order.
+- `source` supplies approved prompt text, text hash, tier and source binding.
+- `identity` contains `model_identity`, `runtime_identity`, `prepared_identity`.
+- `request_sha256`, request settings, encoded tokens, `coverage`, validated artifact hashes.
+- `candidates` contains at most 32 selected neurons with `layer`, `neuron`, `peak`,
+  `peak_position`, `mean_abs`, and full selected per-token `values`.
+- `raw_arrays_in_this_json=false`; original all-neuron reductions use the raw manifest.
+
+`baseline_answer`:
+
+- Same prompt/identity/request join discipline; includes a completed unsteered `answer`,
+  finish reason, usage and final-prefill artifact references.
+- Only revalidated complete source-ledger records qualify. Truncation, failed or uncertain
+  inference remains visible in aggregate source counts and does not enter completed pairs.
+- Final-prefill MLP/residual arrays are not answer-token activation trajectories.
+
+`qualified_labels` and `coherence`:
+
+- `bundle_sha256` binds a reviewed `bundle` of schema
+  `character-creation.public-research-bundle.v1`.
+- Bundle `validation.passed` means **export validation**. Inspect individual evidence
+  and recommendations for scientific pass/fail/unknown outcomes.
+- Label evidence includes all six tested concepts, including failed qualification;
+  original evidence hashes identify retained originals after deployment metadata removal.
+- Coherence evidence includes complete trials, reused-output provenance and failed heldout
+  confirmation. Reuse is not independent replication. Cancelled pilot runs are excluded.
+
+## Measurement semantics: do not collapse these distinctions
+
+1. All-neuron prefill reductions aggregate **all covered encoded prompt positions**, including
+   template tokens. They do not retain the full `[layer, token, neuron]` tensor.
+2. Selected token profiles retain trajectories for the selected neurons only.
+3. Coactivity is an undirected Pearson association. Nearest-neighbor membership can be
+   asymmetric; `source`/`target` storage order is not causal direction.
+4. Weight paths are signed adjacent-layer output/read affinities. They omit attention,
+   normalization and input-conditioned gating; they are not proven functional circuits.
+5. Per-layer PCA axes do not share a physical cross-layer coordinate frame.
+6. A feature may have positive or inverse activation association. That does not determine
+   the dose sign needed to induce behavior.
+7. A visible short path does not prove intervening layers can be skipped. No circuit-memory
+   compiler or accelerated sparse execution is claimed by this release.
+
+## Model binding and steering
+
+Reference checkpoint: `nvidia/Qwen3.8-27B-NVFP4`, revision
+`482ca0f3832238542f8f5295dde86b5f22711d80`; 64 layers, hidden width 5120,
+MLP width 17408. Canonical model metadata is `data/persona_model.json` in Git and
+`persona_model.json` in HF. Recorded model identity:
+`389801c06d68a0d84bd4890b8ce6d26f96e0ee53eb108bcc572c0c09abf5b9a8`.
+
+Neuron addition at the down projection uses:
+
+```text
+requested input addition = dose * median(norm(residual leaving layer L), prompt tokens 1:) / norm(W_down[:, neuron])
+```
+
+Preserve layer/neuron indices, intervention site, phase, calibration hashes, normalized
+dose and actual prepared program identity. Neuron-input addition and post-down column
+addition are separately labelled operations. Quantization and rounding can distinguish them.
+Do not transfer reference Qwen3.5-4B coordinates to the 27B checkpoint. Do not silently
+clamp a fixed recipe or apply a recipe to a mismatching checkpoint.
+
+The public catalog contains 16 authored starting profiles plus six checkpoint-bound
+experimental recipes. Baseline has no pushes. GPT-4o-inspired and Skippy-inspired names
+do not imply original model weights, endorsement, or validated character fidelity.
+
+## Evaluation contracts
+
+Concept qualification freezes candidates on discovery before heldout probes. A compact
+label requires at least 20 distinct predicted-positive heldout families, measured precision,
+a 95% Wilson lower bound **strictly above .70**, balanced accuracy at least .80, recall at
+least .75 and performance above the nuisance baseline by more than .05. Probe-family IDs
+do not establish independent natural examples. Report conditional precision, sample count,
+false positives/negatives, and source hashes—not universal semantic confidence.
+
+Coherence measures supported facts, missing requirements, contradictions and observable
+collapse. A schema representation mismatch can be unknown instead of a factual failure.
+Do not grade optimism, politeness or a preferred personality as coherence. Free prose is
+not certified by a passing fact ledger. Freeze the selected discovery dose before heldout
+confirmation; a heldout failure must not tune a replacement on that same holdout. Recovery
+operates between full responses, not a proven live-token controller. Interaction tests are
+exploratory until independently confirmed.
+
+## Operating a fork
+
+- CPU dependencies: `python -m pip install -r requirements-persona.txt` (includes optional
+  local NLP tools and a hash-pinned spaCy language model). Analysis-only users can install
+  just `datasets huggingface_hub numpy`. Pytest is an additional test dependency.
+- Native model execution requires the separately pinned
+  [Flare runtime](https://github.com/Atlas3DSS/flare-local-inference); CPU dependency install
+  does not prepare GPU graphs or download Qwen. Check local model caches before downloads.
+- Configure admitted endpoints and model/calibration bindings before launching
+  `python persona_explorer.py --config persona_explorer.local.json --port 8792`.
+- Explorer config accepts `nlp_python` (default current interpreter) and
+  `overnight_remote: {ssh_host, hostname, remote_root}`. Set your own values; reference
+  defaults describe the original fleet, not resources a fork may use automatically.
+- Activation launcher accepts `--config`, `--jobs-config` and optional fleet keys
+  `remote_project_root`, `heliopause_path`, `model_path`, `calibration_dir`.
+- Reference multi-card schedulers, GPU leases, thermal checks and memory admission require
+  adaptation to your infrastructure. Single-card/multi-request behavior is not qualified by
+  the six independent B1 workers. Never remove locks to claim concurrent request safety.
+- Detailed setup: [Explorer guide](https://github.com/Atlas3DSS/Character-Creation/blob/main/PERSONA_EXPLORER.md),
+  [porting guide](https://github.com/Atlas3DSS/Character-Creation/blob/main/fork_agent.md).
+
+## Publication lifecycle
+
+`persona/persona_hub.py` manages the approved synthetic prompt subset.
+`persona/persona_public_bundles.py` prepares source-bound, sanitized evaluation bundles and
+calibration assets; it does not upload them. `persona/persona_hub_research.py` publishes
+verified research batches from read-only capture ledgers.
+
+- Metadata watcher polls every five minutes; commits at 500 accumulated records, after
+  15 minutes, or when its source run reaches a terminal state. Completed reviewed bundles
+  accompany the metadata publication.
+- Raw watcher runs on the capture host and uploads up to 32 original NPZ records per
+  immutable tar shard directly to HF. It continues draining the backlog after capture ends.
+- Both verify the pinned remote manifest and uploaded file hashes before advancing durable
+  checkpoints. An uncertain commit reconciles its immutable intent; it never reruns inference.
+- Metadata and raw counts are independent. Read `research_raw_manifest.json` for verified
+  array coverage even if the dataset card or metadata snapshot has an earlier count.
+- Run without `--publish` to validate policy/source files only. Publication requires an
+  explicit local public-release policy binding the exact synthetic corpus and reviewed
+  authored sources. Do not replace those checks with a recursive directory upload.
+- Credentials, private books, book-derived token IDs/profiles, arbitrary interactive prompts,
+  local configs and proprietary research stay private. Book tiers may contribute aggregate
+  progress counts only. Unpublished/failed records remain in the local audit trail.
+
+Core CPU validation uses `tests/test_persona_*`; native CUDA qualification is separate and
+requires an admitted GPU/runtime. NLP tests require the pinned spaCy model. Test skipping
+for absent CUDA is not GPU qualification.
+
+## Attribution and license boundaries
+
+Credit [Lyra / neuron-explorer](https://github.com/lyramakesmusic/neuron-explorer) for the
+foundational steering workflow; this is an independent implementation. Runtime lineage
+and other credits are recorded in the source guide and Flare repository. Synthetic prompts
+are CC BY 4.0. Do not extend that license claim to the upstream checkpoint, derived model
+columns, third-party tools, or private books. Preserve their applicable terms and provenance.

README_human.md

@@ -0,0 +1,76 @@
+# Character Creation: a practical starting point
+
+Character Creation is a workbench for changing an LLM’s behavior through **activation steering**: adding a measured direction or a specific neuron intervention while the model runs. Start with a voice, build a recipe, and compare its answers with an unsteered baseline. A sharper, anxious, playful, or unusually imaginative answer can be an interesting result; sounding pleasant is not the objective.
+
+The approach owes substantial methodological credit to [Lyra’s neuron-explorer](https://github.com/lyramakesmusic/neuron-explorer), especially signed interventions, relative doses, and the distinction between observing neurons and changing outputs. Runtime execution uses the separate [Flare implementation](https://github.com/Atlas3DSS/flare-local-inference). This workbench is an independent implementation, not an endorsement by those projects.
+
+## Start without a GPU
+
+You can inspect prompts, analyze published measurements, and develop evaluation code on an ordinary computer. You do not need to load the model to use the public [Hugging Face dataset](https://huggingface.co/datasets/Atlas3D/character-creation-big-five). The [GitHub repository](https://github.com/Atlas3DSS/Character-Creation) supplies the workbench and analysis code; the Hub repository supplies the public data and published research artifacts.
+
+Install the data tools in your own Python environment:
+
+```bash
+python -m pip install datasets huggingface_hub
+```
+
+Choose a full commit SHA from the dataset’s revision history. Pinning it makes your analysis repeatable even as more evidence is published. Streaming avoids downloading the entire prompt collection:
+
+```python
+from itertools import islice
+from datasets import load_dataset
+
+REVISION = "REPLACE_WITH_A_FULL_HUB_COMMIT_SHA"
+rows = load_dataset(
+    "Atlas3D/character-creation-big-five",
+    name="default", split="train",
+    revision=REVISION, streaming=True,
+)
+for row in islice(rows, 3):
+    print(row["id"], row["prompt"])
+```
+
+The default collection contains synthetic discovery scenarios. Its trait and facet labels describe the prompt-generation assignment, not validated personality scores. The loader calls this split `train`; that does not turn it into an independently held-out evaluation set.
+
+## Know which evidence you downloaded
+
+Read `research_manifest.json` at your pinned revision for published record counts, immutable batches, IDs, and hashes. Read `research_raw_manifest.json` separately for verified raw-array shard coverage. These counts can differ: a metadata record does not prove its raw arrays have finished uploading. Follow the listed shard and entry manifests rather than assuming a directory is complete. Check hashes before joining artifacts to prompts.
+
+There are three important distinctions:
+
+- **Prefill activation search** measures the encoded prompt, including chat-template tokens. It reduces observations across all covered MLP neurons and prompt positions into each neuron’s signed peak, peak position, and sum of absolute activations. This is **not a stored all-neuron-by-all-token matrix**.
+- **Selected token profiles** retain a small selected set of neurons’ values at individual prompt positions. These profiles explain selected candidates; they do not supply trajectories for every neuron.
+- **Full-answer comparisons** contain generated responses and completion status. Baseline answer captures retain final-prefill arrays, not every neuron’s activation throughout the answer. A prefill-only observation cannot substitute for a completed response comparison.
+
+Use coverage and completion fields, not the size of the prompt collection, to describe what was actually measured. Retain failed, interrupted, or truncated outcomes in your analysis instead of quietly counting them as successful answers.
+
+## Run the workbench with your model
+
+Actual steering requires a compatible checkpoint, its calibrated assets, the appropriate Flare environment, and configured inference endpoints. Follow [the Explorer guide](https://github.com/Atlas3DSS/Character-Creation/blob/main/PERSONA_EXPLORER.md) for setup and [the porting guide](https://github.com/Atlas3DSS/Character-Creation/blob/main/fork_agent.md) for another machine. There is no promise that a repository clone alone prepares a model.
+
+Use the Hugging Face `hf` CLI to download the specific calibration assets listed for your chosen publication revision. For example, replace both placeholders with reviewed values:
+
+```bash
+hf download Atlas3D/character-creation-big-five ASSET_FILENAME \
+  --type dataset --revision FULL_COMMIT_SHA --local-dir calibration
+```
+
+Calibration must match the **same checkpoint identity**, including the relevant revision, quantization and runtime preparation. Renaming a checkpoint does not make another model’s calibration compatible. Missing compatible assets mean you must calibrate your own checkpoint; do not bypass identity checks.
+
+After configuring your endpoints and local settings, start the repository’s app wrapper:
+
+```bash
+python persona_explorer.py --config persona_explorer.local.json --port 8792
+```
+
+Open `http://127.0.0.1:8792`. Use the compact workspace picker to move between Persona Studio, recipe experiments, activation search, and evidence views. In the map, **View controls** exposes geometry and connection options; **Expand map** gives the visualization more room. Those display controls do not run inference.
+
+Build a candidate recipe, then compare it against baseline using the same prompts, seed, and answer budget. Check several scenarios before describing a repeatable effect. Distinguish association, weight-based suggestions, and tested interventions.
+
+## Read results with the right limits
+
+A calibrated concept label requires held-out evidence, including a **95% Wilson precision lower bound above 70%**, plus the qualification’s recall, coverage, and balanced-accuracy checks. That is conditional performance on authored text probes—not a universal probability that a neuron “means” a word. Discovery examples select candidates; held-out examples evaluate frozen choices. Neither activation magnitude nor label direction establishes the sign of a useful steering intervention.
+
+Coherence checks target supported facts, missing required details, contradictions, incomplete outputs, and repetition. They do not reward a positive personality. Passing a structured fact ledger does not certify all accompanying prose; fiction may intentionally introduce details.
+
+Private books, extracted passages, arbitrary user prompts, credentials, and local deployment settings are not shipped. Public prompts are licensed **CC BY 4.0**; retain attribution. Derived checkpoint assets retain applicable upstream terms—publishing a measurement does not relicense the underlying model. Consult the manifests and asset notices before redistribution.

layer_00.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d06d7f3f173268f0367a571175ee3e15ada6493610e528c05ee192483f5d1a5
+size 205470

layer_01.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9eeff380e2adbb54b113a15aa24420809bd571c02e0c1a3cfa75b97d2567a29e
+size 192003

layer_02.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee037aca87bd75b55c3079ce6ae358fe34e7aa0364a9eec73592bad00f48910f
+size 198335

layer_03.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:504fd37c56553a0d2f5949a111e56f402326c5166f397b567c8e8ae0eac455c6
+size 204509

layer_04.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:843d9498d2ecb487c8405b43e0f055de38c2db74eba12b26617fe78cce2ff608
+size 218402

layer_05.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b12375e36ecb461735027c24f9bd03850a2ca92fc003d184a330171b26ef5911
+size 204565

layer_06.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0d08a7632420cc6266d0d90dc48eaf5e12264f88dc80c4362dc901c99ab9b53
+size 257435

layer_07.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16a24f559b75b2f36d132fa76efd7ac51b22f9db7a4b18a16da2b4032ba286a5
+size 273722

layer_08.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:718196e419bd3828ed21fc2a0928669c730835f4420c11c97d8385a58709ca1a
+size 262472

layer_09.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8927b22a049f9c900e5d032ec420f2c313b633260c13e687e9c17d8859a1e2cf
+size 247914

layer_10.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7c893ca4320055487d473605f74d00a395cd0c70a2d5c9e26e8a99aed083269
+size 301730

layer_11.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97aa85a347acf66e93d0348c35701e64f955cd4b0a218f2a19322a64a7d89f21
+size 261907

layer_12.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7b25837e22e74bb5b3a2ebd4d5d16a543305a33b9d27b111c11af18571d2c0a
+size 245195

layer_13.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70c9b56987c4e8cd32e047500ad382ca441b858273a19b94da7027bbf922ea93
+size 197653

layer_14.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbc8af4e502a1911a6758a877c93b92e85e8bb2991c84a009ad5f053ccfa468f
+size 250092

layer_15.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76caf39f0f9427ec90122161b8bf61bbc3475b7c9eca7aa59199bf1d5b63241c
+size 261301

layer_16.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02c1c2b6acaa6e1da5337d483e47ec5fbffabbc7434a4fb7b3b4fef2c1130a57
+size 274984

layer_17.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75bc77ff03dcef39081526977ed4af6d4bded79cf8d12990c342c7f583b5a4dd
+size 240783

layer_18.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a48d6d7779fbfb23fa32c90c4ae36c6f1c9cb1c87c9bd578f7a0dab16ac86e7
+size 267758

layer_19.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71c5a8a6566704c6d1fdafb61c3bdca5e8d2a23903b5bbdf6b11f01900de9454
+size 255654

layer_20.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:354549f996a504485af5cc695f72aac02b923ba974b1be3a97ec1f6ef43eefce
+size 236642

layer_21.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9624a6e3ec45c759c794e4ae9c44997644fc8e58f2033dc61bdbe1b927ce338a
+size 257016

layer_22.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9930af0326e2f364c02f1567dfbc86118df46fee0b888cf7a7f3decd36982d79
+size 208656

layer_23.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19e19593e56695b94b12026f47b77f8575ddb3d31289a14e61faf30b56fb168c
+size 245629

layer_24.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5221492e408721a13dc27e18c491c5242fab5ca3ca4ad66d764de5e4a8392774
+size 212521

layer_25.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c09f754162e38d650fb9c8a0e37095f5e76e35f2b75aec8ec4683a97e73be39e
+size 246272

layer_26.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:440909fba92e57bc5f2a186af560b1e299de931005beaf846cbfa8376f448582
+size 244793

layer_27.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d51f73ac2cf6bec55832a295a1c5ec9119e7b5143b2829904bebfa0a49a4c46b
+size 280182

layer_28.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a321eb6a3a0a945111a901312d7511e5f90682bb9352fca51cb1520d1e362e13
+size 231725

layer_29.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e77b4251f3007aebaf8f05539554bb11344687d9c1ad3868b2152a1b975be85
+size 251835

layer_30.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4adb9ecf561a61f8873f91fa3c28c47d9ae43af48dde2bda71b7c82d6df3f3d3
+size 238321

layer_31.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f88b6e7ea450c7caf1b04da8841c79f324e372bc35b52cbb5d4a80586ccc0ba
+size 266427

layer_32.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:640d6ef9bc782b7a100b16d068c9cf025e143f6976008167cf448990c24af4b6
+size 260431

layer_33.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffa190f2bc44542e723d40f4ae604aea8ef290263f0d67918e59a8b476948d19
+size 200796

layer_34.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a35a9274294a8220046bc3207dd42fef2c4b6cda4d1c7fc974fea2e901115a4
+size 285950

layer_35.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5495a50584bf3dfbae2257bce9409eb90aaadc9e764d4827cf95d3d0a0375a20
+size 296393

layer_36.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:710b80ceb6608c1166d29abde70d6c8c78f251956216c8f5c53e4b776cadc4e2
+size 275931

layer_37.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8544456397cc8cd29ea4fa3f4faef12cec431393569e1c6ddd5d9f50f7696367
+size 289767

layer_38.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecc3e8fd5d6c086252499d062beecf580ab025024f80328fab07b4c704b8d601
+size 255768

layer_39.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e57a51401994c3a6c5a59ec02eca13edf7fdd038f45db1eea3d4786bcdfa70a4
+size 296498

layer_40.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c714f3c782720c27be98d81afcd8c0d5451ae830afcea489cdf509c4b7f0c405
+size 275695

layer_41.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40ad167de38f1841855c53fded5d4fb7ce41fe45b4f6b9d892ad6d6fdebdf5f1
+size 256928

layer_42.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8598afafe4f97f324be456b3ad127faa413893d80e7e43318d2029f732ac0b4d
+size 255819

layer_43.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d621643d5611217e0f4bf17aa344be62ced65a379886a5795a1484e585b4db6c
+size 283843

layer_44.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14669a175804ee5dbc640e44f226a1c507a7d3721c6994c437abf2870110a27e
+size 269006

layer_45.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0b6a9e33eaab0b298028c83d89ed713c63a22be8d5090473bc0be8d3618b4bc
+size 240614

layer_46.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54048cc05d15dccfa797a8aa55b24752253c4493c323a9ab38dabdd4267e312e
+size 256747

layer_47.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dfd5ea886a6b1b2074ae403b206b09a68643d89c16c232b964bffb52870ca26
+size 248848