aw-qwen3-8b-v1 — Axiom-World Protocol-v1 Champion (B4v2)

The standalone FP32 export of the final champion adapter of Axiom-World protocol v1: a pre-registered, single-GPU comparison of post-training recipes for rule-grounded planning in a fully verifiable toy world (PlayWorld). This is the two-stage recipe: general-reasoning SFT (GSM8K + MATH-algebra) → PlayWorld task SFT.

• Code & protocol: https://github.com/m97j/axiom-world (tag v1.0.0)
• Tech report: docs/reports/v1/axiom-world-tech-report-v1.md (DOI: 10.5281/zenodo.22052149)
• Run of record: 20260814-023603--b4v2-playworld-sft-from-p1--s42--c56ed2 (full artifacts incl. resolved config & lineage: m97j/aw-runs-b4; 3-seed replications: m97j/aw-runs-seeds)

Model details

The modules_to_save choice is not incidental: attention/MLP-only LoRA on this base model cannot reliably emit the chat template's terminal <|im_end|> (100 % output truncation). Full failure analysis: docs/experiments/v1/adapter_contract_termination.md.

Evaluation (frozen PlayWorld suites, 300 episodes each, greedy decoding)

Pass rate, mean ± sd over training seeds {42, 43, 44}; control = direct task tuning (A2v2) on the identical frozen data:

All 15 suite×seed deltas positive; paired permutation p ≤ 0.0004 each.

Intended use & limitations

This is a research artifact supporting a comparative claim about post-training recipes — not a general-purpose assistant and not a mastered planner (absolute construction-suite pass rates are .30–.39 at this budget/scale; see report §7). PlayWorld is synthetic; transfer to real agent tasks is untested. Outputs are structured-JSON plans for PlayWorld episodes rendered with the bundled chat template.

Release layout and provenance

The repository root contains a standalone FP32 model in standard Transformers safetensors format (5 GB maximum shard size). adapter/ preserves the exact original PEFT weights, config, tokenizer and chat template. The adapter identity digest above combines the adapter config and weight-file digests; it is not the SHA-256 of the safetensors file alone.

provenance/manifest.json records the pinned base and source revisions, library versions, fixed engineering gates, and measured export verification results. Original adapter-only revision: f33d2d16125e89eb38d4b668a2c20a6929ad3784 (tag protocol-v1-adapter). History is preserved. Old root-adapter consumers must pin that revision or use subfolder="adapter". The model includes Qwen weights under Apache-2.0; the Axiom-World code license remains MIT. See the original Qwen model card and license.

The table above reports the original adapter campaign, including three training seeds; this public artifact is the seed-42 champion. Export checks are numerical and standalone-loading checks, not rerun benchmark scores or an ARC result. Greedy generation stops on end-of-text or <|im_end|>; apply the saved chat template for instruction inputs. External benchmark prompting is a separate choice to record with its results.

Precision and verification scope

This export promotes the BF16-loaded base and original adapter to FP32 before merging, and preserves FP32 through saving and standalone reload. It is a precision variant of the historical inference representation. The original campaign scores above are NOT measurements of this FP32 export. Full historical benchmark equivalence has not been established. Numerical merge correctness is checked against the FP32 adapter; comparison with BF16-base PEFT inference is reported separately in provenance/manifest.json under verification.precision.

BF16 export attempts failed the engineering comparison. Load this checkpoint explicitly in FP32; casting to BF16 or quantizing changes the verified execution contract and requires separate evaluation. Benchmark runners should record the checkpoint revision, actual dtype and attention implementation.

How to load

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "m97j/aw-qwen3-8b-v1"
# Pin the published commit from published.json for repeatable benchmarking.
model = AutoModelForCausalLM.from_pretrained(repo, dtype="float32", attn_implementation="sdpa")
tok = AutoTokenizer.from_pretrained(repo)

To inspect the original training representation:

from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B-Base",
    revision="49e3418fbbbca6ecbdf9608b4d22e5a407081db4",
    torch_dtype="bfloat16", attn_implementation="sdpa",
)
model = PeftModel.from_pretrained(base, repo, subfolder="adapter")

For verifier-scored evaluation on the frozen suites, use the code repository's scripts/common/run_evaluation.py with the explicit adapter path and frozen data. See notebooks/protocol_v1/README.md for historical replay boundaries.

Citation

@techreport{kim2026axiomworld,
  author = {Kim, Minjae},
  title  = {Axiom-World: A Pre-Registered Study of Two-Stage Post-Training
            for Rule-Grounded Planning in a Verifiable Toy World},
  year   = {2026},
  doi    = {10.5281/zenodo.22052149},
  note   = {Technical report v1.0. Code: https://github.com/m97j/axiom-world}
}

Post-export task audit and deployment decision

The selected standalone export is FP32. Use explicit dtype=torch.float32 (or the runner's equivalent); automatic downcasting does not reproduce this audit. Both runs used the frozen 1500 tasks, but batch sizes differed: FP32=4, BF16=256. Each row below compares against that run's own adapter control. These are not matched cross-precision results or historical runtime reproduction.

Counts are out of 300 per suite. FP32 is not equivalent to the original adapter: suite-level regressions remain disclosed above. BF16 was rejected for observed within-run degradation, including format failures. This deployment decision followed the results and is not a new preregistered scientific claim. Full paired predictions, verifier verdicts, requests and the decision are in evaluation/precision_audit. Original report metrics describe the historical adapter experiments, not this FP32 export.