Qwen3.5-4B-MATH-ReAct-Agentic-ESOpt

Qwen3.5-4B-MATH-ReAct-Agentic-ESOpt

This is the 25-update Math checkpoint from the Agentic ESOpt run.

Important evaluation note

The reported scores are produced by an agentic, multi-turn ReAct evaluation with a bash tool. They are not one-shot chat or plain transformers.generate() scores. Reproducing the report requires the system prompt, action parser, bash execution loop, answer parser, sampling parameters, and datasets in zz1358m/Agentic-ESOpt.

This checkpoint and the stored Agentic ESOpt report are no-skill (--skill-file ''). The fixed ReAct system prompt is part of the evaluation harness; it is not an external Trace2Skill skill and no skill is baked into the weights.

Weight lineage and integrity

• Base runtime checkpoint: Qwen3.5-4B-text (Qwen/Qwen3.5-4B architecture)
• ES scope: full model, 330 tensors / 4,205,751,296 parameters
• Applied updates: 25, generations 0 through 24
• Update rule: stored ES seeds and z-score-normalized weights, alpha 0.0005
• Weight format: two indexed safetensors shards
• Shard 1 SHA-256: 4563e49054b3d71920bea9c302e350d3fac9184f0c5d19e823c2f1c3183c7283
• Shard 2 SHA-256: 5aee21fccbdca7441612716b6e10abe7bce08763880acb3645e21ec0abc3f9fd

replay_metadata.json records the lineage and replayed_history.json contains the exact 25 update records.

Direct post-export verification

On 2026-08-11, the uploaded checkpoint was loaded directly into vLLM without replaying ES history and evaluated with the report's ReAct/bash protocol, data, one-sample seed, 50-turn limit, 4096-token turn limit, and exact-fallback scorer:

• DAPO100: 0.82 (82/100)
• AIME2026-30: 0.7667 (23/30)
• Request/parser errors: 0
• Final-answer termination: 130/130

This direct check exceeds the stored generation-25 report point (DAPO 0.76, AIME 0.70) and confirms that the exported weights reproduce the reported capability. Sampling is not byte-identical across vLLM engine scheduling, eager/compiled kernels, or GPU types, so compare aggregate metrics.

The stored four-sample final report is:

• DAPO mean: 0.7675; max@4: 0.86
• AIME mean: 0.708333; max@4: 0.966667

Machine-readable results and hashes are in evaluation_results.json.

Report environment

• Python 3.10.12
• CUDA 12.8
• PyTorch 2.10.0+cu128
• vLLM 0.19.1
• Transformers 4.57.6
• Ray 2.56.1
• 4 x 80 GB GPUs, 8 Ray/vLLM engines (gpu_fraction=0.5)
• gpu_memory_utilization=0.42
• context length 131072, no ReAct context trimming
• 50 turns, 4096 generated tokens per turn
• temperature 1.0, top-p 1.0, top-k 40, presence penalty 2.0
• exact-fallback scoring (the report environment did not have math-verify installed)

The dense Qwen3Next compatibility patch in the repository must be enabled for vLLM 0.19.1. The dedicated repository entry point is:

MATH_REPORT_MODE=interval25 scripts/math/eval_hf_checkpoint_4gpu.sh

For the four-sample report:

MATH_REPORT_MODE=final4 scripts/math/eval_hf_checkpoint_4gpu.sh

Dataset SHA-256:

• dapo_test.jsonl: a0e64c93e7801957f0949ab80f5a26233ecd87a02ad5c4628de2da0692b5c4a2
• aime_2026.jsonl: abc8651f3af75ff59341b9de986fef39b1e909aa1466e3b73ee20ec9b6f7242e

Loading the weights

from transformers import AutoModelForCausalLM, AutoTokenizer

path = "zz1358m/Qwen3.5-4B-MATH-ReAct-Agentic-ESOpt"
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, dtype="auto")

Loading succeeds, but the report scores require the agentic harness described above.