README.md

---
pretty_name: Deterministic Random Models
license: other
size_categories:
- n<1K
tags:
- llama
- gemma4
- gemma3
- qwen3
- smollm3
- transformers
- safetensors
- gguf
- synthetic
- conformance
- compatibility-testing
---

# Deterministic Random Models

This dataset contains eleven small, deterministic language-model fixtures for
model-format, loader, inference, compatibility, and conformance testing. They
are not trained models and must not be used for language-model quality
evaluation.

All weights are synthetic and deterministically generated. No original model
checkpoint weights are included.

## Cases

| Case | Architecture | Parameters | Hugging Face | GGUF | Notable feature |
|---|---|---:|---|---|---|
| `tinyllama-chat` | Llama | 303,744 | F32 | Q4_K_M | GQA, query/KV ratio 8 |
| `smollm2-instruct` | Llama | 46,320 | F32 | Q4_K_M | GQA, query/KV ratio 3 |
| `mobilellama-chat` | Llama | 9,296 | F32 | Q4_K_M | MHA |
| `minicpm5` | Llama | 1,409,664 | F32 | Q4_K_M | explicit head dimension, multiple EOS IDs |
| `deepseek-coder` | Llama | 9,296 | F32 | Q4_K_M | linear RoPE scaling |
| `hermes3-llama31` | Llama | 86,336 | F32 | Q4_K_M | Llama 3 RoPE scaling |
| `livekit-turn-detector` | Llama | 132,336 | F32 | Q4_K_M | explicit head dimension, GQA |
| `gemma4-random-model` | Gemma 4 | 6,036,608 | BF16 | Q4_K | five-local/one-global attention schedule |
| `qwen3-random-model` | Qwen 3 | 508,800 | BF16 | Q4_0 | wide Q projection and Q/K head norms |
| `smollm3-random-model` | SmolLM3 | 4,917,504 | BF16 | Q4_0 | three-RoPE/one-no-RoPE layer schedule |
| `gemma3-random-model` | Gemma 3 | 5,938,176 | BF16 | Q4_K | five-local/one-global attention schedule, EOS 106 |

The seven Llama cases are derived from real Hugging Face configuration files by
a preservation-first shrinker. Gemma 4, Gemma 3, Qwen 3, and SmolLM3 retain
architecture-specific reduced geometries that preserve important
ratios, tensor inventories, and layer schedules observed in locally downloaded
upstream GGUF models. Published case names use `random-model` rather than
`tiny-model` to avoid collision with a separately maintained TinyModel collection.

## Formats and layout

The Llama cases retain the original dataset layout:

```text
<llama-case>/
|-- package/model.safetensors       # canonical F32 weights
|-- gguf/model-Q4_K_M.gguf
|-- tokenizer/
|-- reference/outputs.safetensors
|-- inputs.safetensors
|-- case.json
|-- provenance.json
|-- source-config.json
|-- shrunk-config.json
|-- config-diff.json
`-- validation.json
```

The architecture-specific cases use:

```text
<random-model-case>/
|-- hf-bf16/
|   |-- config.json
|   |-- model.safetensors
|   |-- tokenizer.json
|   `-- tokenizer_config.json
|-- gguf-q4_k/ or gguf-q4_0/
|   |-- <case>-Q4_K.gguf or <case>-Q4_0.gguf
|   `-- quantize.log
|-- reference/
|   |-- inputs.json
|   |-- hf-outputs.safetensors
|   `-- gguf-native.json
|-- CONFIG_DECISION.md
`-- metadata.json
```

`manifest.json` is the machine-readable index of all eleven model packages and
their SHA-256 hashes and sizes.

## Synthetic weights and tokenizers

Weights use the `tlfloat::LCG64` recurrence with multiplier
`6364136223846793005`, increment `1442695040888963407`, and ten warm-up steps.
Each case records its seed and provenance.

The reduced models use deterministic 128-token auxiliary vocabularies. These
tokenizers cover token IDs `0..127` and preserve each case's special-token
semantics, but they do not reproduce the linguistic behavior of the original
tokenizer. Explicit token IDs are the primary numerical-test interface.

## GGUF generation and validation

GGUF files were generated with upstream `ggml-org/llama.cpp` commit
`40b740ad05c531b9d57aca6698c3ed553a9e784c`.

Every retained GGUF was loaded through that revision and exercised with direct
token IDs for prefill, cached decode, logit extraction, finite-value checks, and
repeated-execution checks. The effective EOG token set was checked against the
model EOS semantics. Per-case metadata records the actual tensor-type histogram,
hashes, commands, and informational comparison with the corresponding
Transformers reference.

Q4_K, Q4_K_M, and Q4_0 are lossy formats. Their logits are not required to equal the
F32 or BF16 reference exactly.

## Reproducibility and scope

The Hugging Face weights, configs, and GGUF outputs for Gemma 4, Gemma 3, Qwen 3, and SmolLM3 were
independently regenerated and found byte-identical. The Llama cases retain their
source revisions, source-config hashes, shrink decisions, and generation
provenance in each case directory.

This dataset is not a pretrained-model collection, a model-quality benchmark,
or a reproduction of upstream weights or tokenizers. Source-derived configuration
and metadata files may remain subject to terms of their respective upstream
repositories; consult their recorded provenance before redistribution.

See `REPORT.md`, `GGUF_Q4_K_M_REPORT.json`, and
`ARCHITECTURE_RANDOM_MODELS_REPORT.json` for collection-level summaries.

## History

### 2026-08-12: Gemma 4 random model rebuilt

The first `gemma4-random-model` release used hidden width 128, 1,519,168
parameters, and Q4_0. That version was replaced because its small matrix axes
did not exercise K-quant blocks and its GGUF metadata was not sufficiently close
to the inspected 12B Gemma 4 source GGUF.

The current release uses hidden width 256, FFN width 1024, 6,036,608 parameters,
and llama.cpp's `Q4_K` alias. Its actual tensor histogram contains F32, Q4_K,
and Q6_K, matching the source profile family. It preserves the six-layer
five-sliding/one-full schedule, per-layer KV head array, local/global head-width
ratio, dual RoPE regimes, global shared K/V behavior, complete norm inventory,
layer output scales, tied embeddings, tokenizer special IDs, and applicable
sampling metadata. HF BF16 weights and both HF and GGUF references were
regenerated; the previous Q4_0 Gemma 4 files are not part of this release.