This repository hosts PyroDash-4B-GRPO-Lambda-0.05 — the GRPO Stage 3 checkpoint with efficiency penalty λ = 0.05 (quality-oriented), fine-tuned from Qwen/Qwen3.5-4B via PyroDash-4B-SFT.

Companion models: SFT · GRPO λ=0.6

We propose PyroDash, a token-level dynamic reasoning paradigm for collaborative inference between small and large language models. PyroDash enables the small model to autonomously emit the control token <|llm_offload|> during autoregressive streaming decoding; the collaboration engine then dynamically offloads the local reasoning chain to a large model based on this control signal. This approach requires neither an additional router model nor retraining of the large model, and is naturally compatible with closed-source LLM services.

During training, PyroDash follows a three-stage progressive optimization pipeline: (1) train the control-token embedding layer so the small model acquires basic offloading expressiveness; (2) cold-start the offload capability to establish a collaboration pattern between the small and large models; and (3) apply GRPO (this checkpoint, λ=0.05) that jointly optimizes the dynamic offloading policy with a task-accuracy reward and a large-model call-cost penalty, achieving an adaptive balance between reasoning quality and compute cost.

Three-stage progressive training pipeline

Model Details

Item	Value
Base model	Qwen/Qwen3.5-4B
Init	PyroDash-4B-SFT
Stage	(3) GRPO · λ = 0.05 (quality-oriented)
Control token	`<\|llm_offload\|>`
SFT dataset	EasyHard-24K
GRPO dataset	BytedTsinghua-SIA/DAPO-Math-17k
Expert LLM (train/eval)	GLM-5.2-FP8 (frozen)
Precision	bfloat16

Quick Start

1. Setup

bash

git clone https://github.com/PyroMind-Dynamics/pyroDash.git
cd pyroDash
pip install -r requirements.txt

2. Run evaluation (`evaluation/math_eval.sh`)

Edit placeholders in evaluation/math_eval.sh, then:

bash evaluation/math_eval.sh

The script (1) starts a local vLLM server for the small model on port 8001, (2) runs math_eval.py, and (3) stops vLLM on exit.

Parameters

Variable / flag	Meaning	Example
`MODEL`	Local merged model path (vLLM serve + tokenizer)	`/path/to/your/merged_model`
`--glm-base-url`	OpenAI-compatible API for the large/relay model	`http://your-glm-host:8000/v1`
`--glm-api-key`	API key for that endpoint	`your-glm-api-key`
`--glm-model`	Served model name on the GLM side	`your-glm-model`
`--output-dir`	Per-dataset JSON output directory	`./results_500`
`--datasets`	Benchmarks (space-separated)	`gsm8k minerva olympiad aime2024 aime2025`

Tokenizer must include the special token <|llm_offload|>.

Results

Method	Avg. Acc. (%)	LLM Token Ratio (%)	Avg. LLM Calls	Cost ($)
Qwen3.5-4B	28.36	0.00	0.000	2.26
Qwen3.5-4B (+SFT)	46.25	0.00	0.000	1.32
RouteLLM (~75% GLM-5.2-FP8)	52.74	77.37	0.808	44.62
GlimpRouter (τ=0.9)	54.20	75.11	1.20	31.61
PyroDash (λ=0.1)	55.29	8.19	0.058	4.71
PyroDash (λ=0.6)	54.55	1.90	0.012	1.78
PyroDash (λ=0.05) ← this	64.04	95.34	0.975	39.29
GLM-5.2-FP8	57.68	100.00	1.000	49.36

λ=0.05 is the quality operating point (Avg. Acc. 64.04%, above GLM-5.2-FP8 alone). For low cloud cost (~1.9% LLM tokens), use λ=0.6.

Resources

Resource	Link
Project website	PyroMind-Dynamics.github.io/pyroDash
Code	github.com/PyroMind-Dynamics/pyroDash
SFT dataset (EasyHard-24K)	huggingface.co/datasets/pyromind/easyhard-24k
GRPO dataset (DAPO-Math-17k)	huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k
Hugging Face org	huggingface.co/pyromind

Citation

bibtex

@misc{pyrodash2026,
  title        = {PyroDash: Cost-Efficient Token-Level Small-Large Model Collaborative Inference},
  author       = {{PyroMind Dynamics}},
  year         = {2026},
  note         = {Preprint}
}

@misc{pyromind2026easyhard24k,
  title        = {{EasyHard-24K} v0.02},
  author       = {{PyroMind Dynamics}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/pyromind/easyhard-24k}}
}

GRPO training data: BytedTsinghua-SIA/DAPO-Math-17k.

License

Apache 2.0 (derived from Qwen3.5-4B).

pyromind/PyroDash-4B-GRPO-Lambda-0.05

PyroDash-4B-GRPO-Lambda-0.05