Publishing TinyLiquid Analyst on Hugging Face (for downloads → grants/loans)

1. Prepare the artifact

export PYTHONPATH=$PWD
# after the training pipeline finishes (ckpt/dpo exists):
.venv/bin/python hf/export_hf.py --ckpt ckpt/dpo --out hf_repo     # safetensors + q8 + configs + modeling file
.venv/bin/python hf/export_gguf.py --ckpt ckpt/dpo --out hf_repo/tiny-liquid-q8.gguf
.venv/bin/python eval/bench.py --ckpt ckpt/dpo --out bench/metrics.json
.venv/bin/python hf/build_card.py --metrics bench/metrics.json     # model card from real metrics

Sanity checks before publishing:

# 1) transformers path (trust_remote_code) produces coherent text
.venv/bin/python - << 'PY'
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("hf_repo")
m = AutoModelForCausalLM.from_pretrained("hf_repo", trust_remote_code=True)
m.persona_id = 1
ids = tok("<|analyst|><|user|>Verify: 'The bridge was painted in 2019 and never repainted.' Records show a 2022 repaint permit.<|assistant|>", return_tensors="pt").input_ids
print(tok.decode(m.generate(ids, max_new_tokens=80, do_sample=True)[0]))
PY

# 2) GGUF round-trips natively
.venv/bin/python -m model.gguf_runtime --gguf hf_repo/tiny-liquid-q8.gguf \
    --prompt "<|analyst|><|user|>What's your take on coincidences?<|assistant|>"

2. Publish

huggingface-cli login          # paste your HF token
.venv/bin/python hf_upload.py --repo YOURNAME/tiny-liquid-analyst

Then on the HF web page:

set the License to apache-2.0 (already in the model card metadata),
add tags: tiny-model, on-device, liquid-architecture, fact-checking, osint, gguf,
add a demo (optional): demo/serve.py behind a tunnel, or the HF Spaces template,
pin the README's YOUR-ORG links after upload.

3. Grant/loan-ready framing (what reviewers look for)

Reproducibility: every step in this repo is scripted (run_*.sh, data/gen_*.py, hf/export_*.py, eval/bench.py). Include the commit hash in your application.
Originality: non-transformer liquid architecture, own tokenizer, own data pipeline, own SOP/agent tooling — nothing is a wrapper around another model.
Efficiency story: 7.8M params, trained on an 8-core ARM laptop with no GPU, quantized to ~5-11 MB. That is the headline for edge-AI grants: SOTA-scale capability per watt.
Evidence: bench/metrics.json (val perplexity, probe accuracy, tok/s) plus generation samples in the model card. Add a short technical report citing them.
Guardrails: the OSINT/dark-web tooling is scoped to authorized research with hard stop rules — show this explicitly; it de-risks your application.
Community: answer questions on the HF discussion tab, add a Spaces demo, and post quantization/config updates. Downloads follow usefulness, not hype.

4. Growth levers after v1

Code stage (continuation pretraining on data/code_train.bin) — planned next.
Bigger variant (e.g., 30-60M MoE) once a GPU or cloud budget appears.
Multilingual tokenizer + a second persona language.
ONNX export + onnxruntime for even faster ARM inference.