Rewrite dataset card for default and dense-100m

Clarify configuration choice, pretraining/SFT/SDFT usage, dense schema, token scale, and integrity guarantees.

README.md

@@ -7,6 +7,7 @@ language:
 tags:
 - synthetic
 - state-space-models
+- recurrent-models
 - pretraining
 - sft
 - sdft
@@ -31,77 +32,99 @@ configs:
 
 # StateBench v1
 
-### 10 million executable state programs for memory, recurrence, and state control.
+### 110 million verified state episodes for pretraining, SFT, and on-policy SDFT
 
-One deterministic corpus. Three training views: **pretraining · SFT · on-policy SDFT**.
+Train models to **retain**, **edit**, **address**, and **transform** state across long execution traces.
 
 </div>
 
-| Scale | Coverage | Storage | Splits |
-|:--|:--|:--|:--|
-| **10,000,000 rows** | **36 families · M1–M11** | **7.80 GiB · 41 Parquet shards** | **9,949,600 train · 50,400 validation** |
+## Choose a configuration
 
-| Dense extension | Physical rows | Density | Storage |
+| Configuration | Best for | Scale | Representation |
 |:--|:--|:--|:--|
-| **100,000,000 episodes** | **25,000,000 packed rows** | **162.01 units/1k chars** | **94.78 GiB · 251 shards** |
+| **default** | Broad behavioral coverage and varied surfaces | 10,000,000 programs · 7.80 GiB | 36 families across M1–M11 |
+| **dense-100m** | High-throughput state-intensive training | 100M episodes · 94.78 GiB | 4 episodes/row · compact shared VM |
 
-StateBench teaches models to execute finite state programs rather than recall facts. A typed program is executed by a deterministic reference evaluator; only then is it rendered into varied textual surfaces.
+Both configurations contain structured prompt, completion, and teacher-only privileged context. Every target is computed by an executable reference machine, never authored by a teacher model.
 
-## Load it
+## Load
 
     from datasets import load_dataset
 
-    ds = load_dataset("aabbdev/StateBench-v1", split="train", streaming=True)
-    dense = load_dataset("aabbdev/StateBench-v1", "dense-100m", split="train", streaming=True)
+    default = load_dataset(
+        "aabbdev/StateBench-v1", "default", split="train", streaming=True
+    )
+    dense = load_dataset(
+        "aabbdev/StateBench-v1", "dense-100m", split="train", streaming=True
+    )
 
-    pretrain = ds.select_columns(["text"])
-    sft = ds.select_columns(["prompt", "completion"])
-    sdft = ds.select_columns(["prompt", "privileged_context"])
+| Split | default | dense-100m |
+|:--|--:|--:|
+| Train | 9,949,600 rows | 24,875,000 rows / 99.5M episodes |
+| Validation | 50,400 rows | 125,000 rows / 0.5M episodes |
 
-The default configuration stores text directly. The dense-100m configuration omits that
-redundant physical column; derive it from prompt and completion when using causal pretraining.
+## Train three ways
 
-## One corpus, three views
+### SFT
 
-| Regime | Student input | Training signal |
-|:--|:--|:--|
-| Pretraining | text | Complete causal document |
-| SFT | prompt | completion |
-| SDFT | prompt | privileged_context, visible only to the teacher |
+    sft = dense.select_columns(["prompt", "completion"])
 
-> **SDFT contract:** privileged_context contains the verified rendered target completion. Never concatenate it into the student prompt. It is designed for TRL SDFTTrainer teacher conditioning.
+### On-policy SDFT with TRL
 
-## What is inside
+    sdft = dense.select_columns(["prompt", "privileged_context"])
 
-| Capability | Families |
-|:--|:--|
-| Retain and address | passive retention, capacity, MQAR, multi-token keys, collision resistance |
-| Edit and select | overwrite, erase/reuse, masked edits, selective write/copy, scoped reset |
-| Track dynamics | DFA, Mealy, permutations, monoids, modular counters, finite groups |
-| Adapt online | one-shot binding, remapping, rule switching |
-| Survive long execution | active recurrence, cycles, checkpoints, read/generation pollution |
+privileged_context is visible only to the teacher. Never concatenate it into the student prompt. It contains the verified rendered target expected by TRL SDFTTrainer.
+
+### Causal pretraining
+
+The default configuration includes a ready-to-use text column. dense-100m avoids storing a duplicate ~95 GiB text view; derive it while streaming:
+
+    def to_text(row):
+        prompt = row["prompt"][0]["content"]
+        answer = row["completion"][0]["content"]
+        return {"text": f"User:\n{prompt}\n\nAssistant:\n{answer}"}
 
-Difficulty is balanced across levels 0, 2, 4, 5, and 6. Family allocation differs by at most one row across the full corpus.
+    dense_pretrain = dense.map(to_text)
 
-## Canonical columns
+## What models practice
 
-| Column | Meaning |
+| Capability | Examples |
 |:--|:--|
-| prompt | Conversational student-facing input |
-| completion | Verified assistant target for SFT |
-| privileged_context | Teacher-only verified solution for on-policy SDFT |
-| text | Prompt and completion serialized for causal pretraining |
-| semantic_fingerprint | Deduplication and semantic identity |
-| family, macro, level | Curriculum controls |
-| surface_json, transformation_ids | Rendering provenance |
+| Retention and capacity | Delayed recall, bounded buffers, long-lived registers |
+| Addressing and editing | Multi-query recall, collisions, overwrite, erase/reuse |
+| Selection and interference | Gated writes, scoped memory, proactive and retroactive interference |
+| State dynamics | Automata, counters, permutations, monoids, finite groups |
+| Online adaptation | One-shot binding, remapping, rule switching |
+| Long execution stability | Active recurrence, checkpoints, read and generation pollution |
+
+default implements 36 explicit behavioral families. dense-100m maps the same curriculum labels onto a compact shared register VM for maximum state-transition density; it is a dense training curriculum, not a byte-for-byte replica of default prompts.
+
+## Data contract
+
+| Field | default | dense-100m | Purpose |
+|:--|:--:|:--:|:--|
+| prompt | yes | yes | Student-visible conversational input |
+| completion | yes | yes | Verified SFT target |
+| privileged_context | yes | yes | Teacher-only SDFT context |
+| text | yes | derived | Causal pretraining view |
+| semantic identity | one fingerprint | four fingerprints | Deduplication and audit |
+| family / macro / level | scalar | four-item lists | Curriculum filtering |
+| density_units_per_1k_chars | no | yes | Semantic density control |
+
+## dense-100m at a glance
+
+| Episodes | Physical rows | Density | Token estimate | Generator |
+|--:|--:|--:|--:|:--|
+| 100,000,000 | 25,000,000 | 162.01 units / 1k chars | 107–120B tokens | C++20 + Arrow |
 
-## Trust and scope
+The token range was measured on 1,000 real packed rows with GPT-2 and Qwen3 tokenizers. Actual training cost depends on chat templates, truncation, packing, and the SDFT generation policy.
 
-- Every oracle is produced by the reference evaluator, not written by a teacher model.
-- Programs are deduplicated by semantic fingerprint.
-- dataset_manifest.json records generation parameters, distributions, shard sizes, and SHA-256 checksums.
-- The public corpus is training data, not the private StateBench evaluation split.
+## Integrity
 
-Generated with **StateBench-v1.1**. The canonical dataset contains both SFT and SDFT fields so no semantic example is duplicated between training formats.
+- 110M semantic identities are unique across their respective releases.
+- Required fields, roles, answer JSON, and SDFT isolation passed full-corpus audits.
+- No benchmark name, family label, or privileged answer appears in model-visible prompts.
+- SHA-256 checksums cover every Parquet shard.
+- The public training corpus is separate from private StateBench evaluation material.
 
-The v1.1 data revision removes benchmark/family labels from model-visible prompts, emits one valid JSON output template for multi-field answers, and omits empty semantic fields. For the longest examples, configure the trainer context explicitly; the provided SFT/SDFT scripts default to conservative 17,408/16,384-token limits.
+Release manifests: dataset_manifest.json for default and dense-100m-manifest.json for dense-100m.