zz1358m/Qwen3-30B-A3B-Agentic-ESOpt-DocVQA

Qwen3-30B-A3B Agentic ESOpt DocVQA (theta40, exact TP=2 state)

This repository contains the exact tensor-parallel state produced by the 40-update full-parameter Agentic-ESOpt DocVQA 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 best broadly performing one of the four consecutive four-sample groups in that Eval16 run, with sample_index 4–7. ANLS Mean4 is the mean ANLS over those samples; ANLS Pass4 is the mean per-problem maximum ANLS. Accuracy Mean4 is exact-match accuracy over all four samples; Accuracy Pass4 is the fraction of problems with at least one exact match. These remain results from the original Eval16; the later four-sample reload check validates checkpoint equivalence 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-DocVQA",
    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 that exact state cannot be represented by one ordinary unsharded Hugging Face state dict.

Training/replay setting

• —Base model: Qwen3-30B-A3B
• —Updates: 40
• —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)
• —DocVQA environment: paper-aligned multi-turn CLI/OCR 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/docvqa_eval16.json: original 16-sample evaluation records
• —repro/metrics.json: original reported Eval16 metrics and aggregation
• —repro/es_history_theta40.json: exact history used for reconstruction
• —repro/aligned_history_replay_eval4.json: aligned four-sample history replay
• —repro/reloaded_sharded_eval4.json: aligned four-sample sharded reload check
• —repro/verification.json: reload-equivalence comparison
• —repro/SHA256SUMS: published-file checksums