-- ▄▖▄▖▄▖▄▖▄ -- ▛▖▞▌▐ ▌ ▙▘▌▌▌▌ -- ▌▝ ▌▟▖▙▖▌▌▙▌▙▘ v2.5

MICROD-BIGRAM-V2.5

This model was trained from scratch using a combination of:

• Simplified policy-gradient (GRPO-style) for character-level trajectory training
• VAE fields recorded as metadata only (do not affect the transition table)
• 4× quantum harmonic oscillator superposition for deterministic seeding

It is a compact character-level bigram model — no transformer weights, just transition probabilities. Runs entirely in the browser or on any machine with Python 3.8+.

This model was produced with Micro Distillery v2.5 — a single-file HTML PWA that fetches open corpora, trains a bigram model, applies optional NLP fan-out polish, and exports a HuggingFace-ready .zip bundle. Includes the Jev Code Nous deterministic API layer for coding-focused token prediction (harness only). Available for download in the repo folder. Here

CHANGE LOG — Updated 2026-10-25

Maintenance + feature release for the browser trainer. Bundle format remains mD-v2 / 2.5.0. No breaking changes to the exported artifact schema.

Added

• Enhanced Fan-out NLP Polish (v2) — configurable LanguageTool (off / public / custom URL), MyMemory paraphrase round-trip with daily quota tracking, Datamuse lexical scoring + cache. Status badges, diagnostics table, and in-flight guards.
• Jev Code — Nous Deterministic API Layer — harness-only coding-focused token prediction pipeline (tokenize → context → candidates → score → select → validate). Local, seeded, no network. Per-request yes/no confirmation modal.
• Self-host LanguageTool quick-start — Docker and MindTab proxy snippets in the UI.
• QFM (Quality Filter Module) inspect for harness uploads and artifact validation.
• Export options — selective inclusion of config, vocab, logs, knowledge graph, diagrams, and to_parquet.py.

Fixed (fan-out review pass N01–N25)

• N01/N02: ltStatus gains 'unverified' state; badges render ? for it.
• N03/N24: MyMemory quota-exhausted is distinct ('quota') and pre-checked with round-trip headroom.
• N06/N07: HTTP-prefixed errors classify as 'unreachable', not 'cors-blocked'.
• N08: LanguageTool test button disabled when mode is off.
• N09: Fan-out “Run test” has an in-flight guard.
• N12: lt-status-label consistent with mode (shows mode · ? when unverified).
• N14: Cache column distinguishes 'partial'.
• N16: LT test reads live URL input value.
• N18: Custom URL auto-prepends http:// when scheme is missing.
• N20: Truncated MyMemory q is trimmed to word boundary.
• N21: Wizard-seen localStorage flag set only after openWizard succeeds.
• N25: Fan-out diagnostics message cap raised to 48 with ellipsis.

Changed

• Version string v2.5 / format mD-v2 / artifact version 2.5.0.
• Honest training notes expanded in exported README.md and config.json (no real transformer, no real VAE, no emitted EOS token, heuristic reward).
• Mutual exclusion of async ops (fetch / train / harness) via state.training + state.fetching.
• LanguageTool defaults to off (public endpoint is CORS-blocked from browsers).

Notes

• This remains a base model and needs added datasets to it's folder to be more robust. The “model” is a 27×27 character-bigram transition table.
• VAE latent_dim and filter_threshold appear in config for reference only and do not affect transitions.
• No explicit EOS token is emitted; generation is bounded by max_tokens. BOS (space) is the practical trajectory end signal.
• Jev Code runs only inside the Harness modal and never makes network calls.

---

Reference Architectures

The bundle metadata references these modern open models as inspiration sources. The actual model is a distilled bigram — not a transformer checkpoint — and does not load via AutoModelForCausalLM.

• SmolLM2-135M (HuggingFaceTB)
• Qwen2.5-0.5B (Qwen)
• OLMo-2-1B (AllenAI)
• TinyLlama-1.1B (TinyLlama)

Model Details

• Format: mD-v2
• Version: 2.5.0
• Reference arch: tiny-bigram (default) — selectable metadata: SmolLM2-135M / Qwen2.5-0.5B / OLMo-2-1B / TinyLlama-1.1B
• Vocab size: 27 tokens (a–z + space)
• Transition states: 27 (source states)
• Harmonic seed: 1251963776 (default; driven by 4× oscillator measurement)
• Training method: simplified policy-gradient with KL mixing
• Created: session timestamp (ISO)

Training Notes

• No transformer, no attention, no GPU. The model is a character-bigram transition table.
• “GRPO” is simplified. Trajectories are sampled, scored with a heuristic text-shape reward (length, alpha-only, vowel-ending, no-triples), advantages are z-scored, high-advantage transitions are multiplied by (1 + lr × advantage), then KL-mixed toward a reference snapshot. Reported loss is -mean(advantage × log_prob).
• VAE fields are metadata-only. They do not filter or encode anything.
• Reward is heuristic, not a learned reward model.
• No explicit EOS token. Sentinels are declared for schema compatibility only and are never emitted.

Hardware Requirements

Component Minimum Requirements Recommended (for Self-Hosting Extensions) Processor (CPU) Any modern CPU (Intel Atom, Celeron, Raspberry Pi 3+, or newer) Intel Core i3 / AMD Ryzen 3 or equivalent Graphics (GPU) None (Not Supported) None (Not Supported) Memory (RAM) 128 MB (Total script footprint is < 5 MB) 4 GB to 8 GB (Only if running optional Docker containers) Storage < 1 MB for the model files 10 GB (If downloading external training datasets)

Test Wrappers Included

cd /the folder you have the files in/

# 1. Run the test harness
chmod +x test.sh run.sh
./test.sh

# 2. Or chat directly
python3 test.py

Optional: Enhance test.sh / run.sh to Automate the Suite

Update your bash test harness script to systematically validate predictions, catch empty states, and verify output stability across your custom files.

Optional: How to Add Data & Enhance the Tests

Because this is an ultra-compact character model trained primarily on Andrej Karpathy’s names.txt corpus (plus optional README cards), you can enhance coverage by feeding alternative datasets via the browser Harness (drag-and-drop .txt / .md / .json / .zip) or by updating the wrapper code after export.

How the Quantum Oscillators Work for Seeding

The 4× quantum harmonic oscillator superposition does not process token prediction step-by-step during inference. It runs at the end of the pipeline to establish a deterministic seed.

Component	Operational Role
Simulated Superposition	Algorithmic simulation of 4 oscillators with frequencies ω = 1, φ, 2, π.
Wavefunction Collapse	Measurement collapses the superposition to a single integer (default harmonic seed 1251963776).
Deterministic Initialization	The seed locks the PRNG used for sampling and reproducible transition probabilities in model.json.

Files in the Bundle

| File                        | Description                                      |
| --------------------------- | ------------------------------------------------ |
| `model.json`                | Trained bigram model (transitions + config)      |
| `config.json`               | HuggingFace-style metadata                       |
| `vocab.json`                | Character vocabulary                             |
| `tokenizer_config.json`     | Tokenizer metadata (sentinels declared only)     |
| `knowledge_graph.json`      | 2D graph seeded from corpus                      |
| `training_logs.json`        | Per-step metrics + quantum measurements          |
| `test.py`                   | Interactive chat wrapper                         |
| `test.sh`                   | Test harness                                     |
| `test_readme.md`            | Testing instructions                             |
| `run.sh`                    | Minimal launcher                                 |
| `to_parquet.py`             | Convert transitions to parquet                   |
| `manifest.json`             | Bundle checksums (self-excluded)                 |
| `diagrams/pipeline.svg`     | Training pipeline diagram                        |
| `diagrams/architecture.svg` | Model architecture diagram                       |
| `README.md`                 | This model card                                  |
| `LICENSE`                   | Apache 2.0                                       |

Training Pipeline

1. Fetch — pull corpus from HuggingFace + GitHub (open 2026 endpoints) or use bundled fallback
2. Knowledge graph — build character-transition graph, laid out in 2D
3. Seed — initialize bigram transition counts from the corpus
4. Simplified GRPO — sample trajectories, heuristic reward, advantage update + KL mix
5. VAE metadata — record latent_dim / filter_threshold (no effect on table)
6. Quantum seed — 4 harmonic oscillators (ω = 1, φ, 2, π) collapse to deterministic seed
7. Export — HuggingFace-ready .zip with model card, tests, diagrams, manifest

Optional at inference time (browser only): Fan-out NLP polish (LanguageTool / MyMemory / Datamuse) and Jev Code deterministic coding token prediction.

Data Sources

This bundle was produced by fetching from these free, open endpoints (with cache fallback):

• raw.githubusercontent.com/karpathy/makemore/master/names.txt (training corpus)
• huggingface.co/HuggingFaceTB/SmolLM2-135M/raw/main/README.md (model card)
• huggingface.co/api/models?search=smollm2&limit=5 (KG enrichment only)
• (Optional / fan-out) api.languagetool.org/v2/check, api.mymemory.translated.net/get, api.datamuse.com/words

License

Apache 2.0