Qwen3-30B-A3B-Agentic-ESOpt-Math

Qwen3-30B-A3B Agentic ESOpt Math (theta20, exact TP=2 state)

This repository contains the exact tensor-parallel state produced by the 20-update full-parameter Agentic-ESOpt Math run.

Checkpoint correction (2026-09-18): The checkpoint files initially uploaded to this repository were exported incorrectly and did not reproduce the trained TP=2 runtime state. The TP=2 vLLM sharded-state checkpoint uploaded on 2026-09-18 is the correct checkpoint.

Reported result

The source evaluation generated 16 samples per problem. The reported Mean4/Pass4 row is the group with sample_index 12–15. Mean4 is the mean score over those four samples; Pass4 is the mean per-problem maximum over those four samples. These remain the original Eval16 results; the later aligned four-sample replay is included as a reproduction artifact and does not replace the reported scores.

Required loading configuration

The checkpoint is a vLLM sharded-state checkpoint and must be loaded with tensor parallel size 2:

from vllm import LLM

llm = LLM(
    model="zz1358m/Qwen3-30B-A3B-Agentic-ESOpt-Math",
    tokenizer="Qwen/Qwen3-30B-A3B",
    load_format="sharded_state",
    tensor_parallel_size=2,
    dtype="bfloat16",
    trust_remote_code=True,
)

Do not load this checkpoint with transformers.AutoModelForCausalLM, TP=1, or a tensor-parallel size other than 2. Full-parameter ES used independent, rank-salted random streams on the two TP workers. Some vLLM parameters are replicated across TP ranks, so the exact state cannot be represented by one ordinary unsharded Hugging Face state dict.

Training/replay setting

• Base model: Qwen3-30B-A3B
• Updates: 20 (generations 0–19)
• Population: 16
• Case batch: 16
• Parameter scope: full
• Tensor parallelism: TP=2
• Sigma: 0.001 to 0.0005, cosine schedule
• Alpha: 0.0005
• Reward normalization: population z-score (ddof=0)
• Math environment: paper-aligned 50-turn CLI agent

es_sharded_export_manifest.json records the source history hash, replayed generation count, required load format, TP size, and emitted worker shards.

Reproduction files

• repro/dapo_eval16.json: original DAPO 16-sample evaluation records
• repro/aime2026_eval16.json: original AIME2026 16-sample records
• repro/metrics.json: original reported Eval16 metrics and aggregation
• repro/es_history_theta20.json: exact history used for reconstruction
• repro/aligned_history_replay_eval4.json: aligned four-sample replay
• repro/SHA256SUMS: published-file checksums