Rizzo Flow 1.7B: Spark-X2.5-1.7B + typed-decisions LoRA

This is a LoRA fine-tune of XHToken/Spark-X2.5-1.7B for Rizzo Flow. It is the small variant: when accuracy matters, use the 4B, rizzoaiacademy/rizzo-flow.

Rizzo Flow turns an unstructured state into typed decisions (boolean, choice, score) with probabilities, and it generates no text. The model runs one forward pass per question. The answer is the softmax over the logits of the answer letters (A, B, …) at the last prompt position. Training works the same way: the loss is a soft cross-entropy between the target distribution and that softmax, and only the LoRA adapters are trained.

This model is not affiliated with TypeSafe, and it is not Jev.

Files

Use

These GGUF files are what Rizzo Flow downloads with --size 1.7b. It pins this repository by commit and checks the sha256 of every file:

git clone https://github.com/Rizzo-AI-Academy/rizzo-flow && cd rizzo-flow
uv sync --locked
uv run rizzo download --size 1.7b   # llama.cpp for this machine + this model, Q8_0 (1.8 GB)
uv run rizzo serve --size 1.7b      # → http://127.0.0.1:8017/playground

--quant bf16 or --quant q4_k_m picks another file and --weights base the original Spark-X2.5 GGUF. The server answers as rizzo-flow-1.7b-q8_0.

MLX (Apple Silicon, or MLX-CUDA) uses the safetensors checkpoint at the root of this repository instead of the GGUF files:

uv sync --locked --extra mlx
uv run --no-sync rizzo download --backend mlx --size 1.7b      # the safetensors only, 3.4 GB
uv run --no-sync rizzo serve --backend mlx --size 1.7b --bits 8

It answers as rizzo-flow-1.7b-bf16 (or …-q8/…-q4 with --bits). On an RTX 5060 Ti with MLX-CUDA, BF16, the 4B gives the same prompts (same hashes) and the same probabilities as llama.cpp BF16 within 0.001.

transformers loads the checkpoint with the original Spark code (trust_remote_code=True, modeling_spark.py is XHToken's file, unchanged): AutoModelForCausalLM.from_pretrained( "rizzoaiacademy/rizzo-flow-1.7b", trust_remote_code=True, dtype="bfloat16"). Use the prompts of Rizzo Flow and read the answer-letter logits at the last position; generating text is not what it was trained for.

The GGUF files are plain GGUF (architecture spark2_5; Rizzo Flow runs them on llama.cpp release b11081), but the model is meant to be read, not chatted with: Rizzo Flow builds the prompt, then reads the logits of the answer letters. In a chat app it behaves like a lightly changed Spark-X2.5, and the fine-tune brings nothing there.

The adapter was trained on the exact prompts of PROMPT_VERSION = "spark-decisions-v3", commit c30cc63 of rizzo-flow. Any other prompt format can make the adapter useless.

Results

We ran the test split of the LocalLLaMA/typed-decisions benchmark (config all, 400 cases, 2,000 decisions). The model is fine-tuned for typed decisions, so this is not a zero-shot result, but training used no data from these four workflows: the Open-Jev config that shares them (workflow-controls-v1) was excluded, and so was every training state containing text from the test set. The rows below come from the same machine (RTX PRO 6000) and the same runtime (llama.cpp b11081, CUDA); on an RTX 5060 Ti we measured 0.530 → 0.544 (paired difference +0.014, 95% interval [−0.014, +0.043]) and KL 3.031 → 0.694:

Fine-tuning mostly improves the probabilities (KL, Brier, ECE). Accuracy barely moves: +0.015 is within noise. The change is not uniform across workflows:

On the dev split of the training mixture (600 questions), loss goes 3.117 → 0.460 and accuracy 0.408 → 0.820.

Q4_K_M costs accuracy on this model. Made with llama-quantize b11081 from the BF16 GGUF (no importance matrix), it scores 0.490 on the RTX 5060 Ti against 0.544 for Q8_0: −0.054, 95% interval [−0.070, −0.037]. KL 0.640, Brier 0.279, ECE 0.192; invoice_processing falls to 0.394 and security_incidents to 0.474. Use it only if 1.8 GB does not fit; the 4B at Q4_K_M (2.6 GB) is much better.

The probabilities are not calibrated. Fit temperature scaling on data from your own domain (rizzo calibrate) before you use thresholds. We have not yet re-run the smoke and SemIf fixtures on this model.

Training

• Base: Spark-X2.5-1.7B BF16, rev. 14d6e83. LoRA r 16, alpha 32, dropout 0.05, on all linear layers of attention and MLP. Embedding and lm_head are not trained because they are tied. 16.7M trainable parameters.
• AdamW, lr 5e-5, 3% warmup then cosine decay, no weight decay, clip 1.0. 16 questions per step. 1 epoch = 1,770 steps. Seed 0. The data and hyperparameters are the same as for the 4B.
• 28,321 training questions (14.7M prompt tokens, max 2,048 tokens per prompt), from tasksource/procedural-typed-decisions (Apache-2.0), ZefanCai/Open-Jev release-v2-redistributable (CC0) and 12 configs of Praveenrajus/jev-bench. None of the labels come from Jev or TypeSafe. We removed every training state that contains text from the evaluation sets.
• Hardware: 1× RTX PRO 6000 Blackwell 96 GB, batch 8 / 3,072 tokens per forward, no gradient checkpointing. Throughput ~15,500 prompt tokens/s, 19 minutes, 22 GiB peak.

Full pipeline and reasoning: docs/training.md.

Licenses of the training data

The base model is Apache-2.0. Among the jev-bench configs, stsb is CC-BY-SA. measuring_hate_speech and the HelpSteer2 configs are CC-BY-4.0: Measuring Hate Speech (Kennedy et al., 2020; Sachdeva et al., 2022) and HelpSteer2 (Wang et al., 2024, NVIDIA). Check whether these terms matter for your use before you redistribute derivatives.