SEDD-medium Math Tool SFT

这是 louaaron/sedd-medium 的全参数监督微调版本。模型学习的目标不是直接 心算答案，而是在引用式协议中逐步产生工具动作；外部精确计算器执行 ADD / SUB / MUL / DIV，模型再引用工具结果给出最终答案。

Problem: 7 + 9 * 7
<CALL>MUL(x1,x2)</CALL>  -> r0=63
<CALL>ADD(x0,r0)</CALL>  -> r1=70
<FINAL>r1</FINAL>

This repository contains the complete online FP32 inference state. It does not contain the 6.79 GB training checkpoint, optimizer, EMA, RNG state, sealed test prompts, or per-example evaluation traces. The upstream base weights are not needed after loading this full-parameter state.

Artifact identity

Item	Value
Architecture	SEDD-medium / DDiT, 24 blocks, hidden size 1024
Parameters	424,477,010 model parameters (+1 learned noise parameter)
Context length	256 tokens
Weight variant	online FP32
Weight tensors	251
model.safetensors	1,697,932,808 bytes
Weight SHA-256	f3ee746f5ab56f8549592eca81bb33198b66ff16940cd6959e16943a5b06ee4a
Source checkpoint SHA-256	020f76c434acb7652adcf0bc09f26c4605a00e7fe4462b6f71cea7f91438ebd5
Reproducible code	Jrffy666/SEDD@9b26b898

Machine-readable provenance, dimensions, file sizes, and hashes are in inference_metadata.json and checksums.sha256.

Loading the weights

This is a custom SEDD architecture, not a Transformers PreTrainedModel. AutoModel.from_pretrained() and the Hugging Face inference widget are not supported. The included loader checks filenames, sizes, SHA-256 digests, configuration schema, tensor keys, shapes, dtypes, and finite values before moving the model to the requested device.

pip install -r requirements.txt
hf download Jrffy666/SEDD-medium --local-dir ./SEDD-medium

import sys
from pathlib import Path

snapshot = Path("./SEDD-medium").resolve()
sys.path.insert(0, str(snapshot))

from hf_model_loader import load_hf_sedd_model

model, graph, noise, config, metadata = load_hf_sedd_model(
    snapshot,
    device="cuda",
)
print(metadata["source"]["training_checkpoint_sha256"])

The strict loader verifies roughly 1.70 GB of data and therefore performs a full SHA-256 pass by default. Pass verify=False only if the surrounding storage layer already provides equivalent integrity verification; structural and dtype checks remain enabled.

For the tokenizer, reference-tool loop, exact Fraction calculator, training pipeline, evaluation scripts, and FastAPI/React demo, use the Jrffy666/SEDD source repository. The model emits protocol actions rather than ordinary chat responses.

Training summary

• Base: louaaron/sedd-medium, pinned upstream revision ce71a3c6178b50e899c8be1a1a4c13130308e54f.
• Data generator: Google DeepMind Mathematics Dataset, pinned commit 427f45075f84b8b9774950196ad63867ca20ffb3.
• Objective: reference-tool action SFT with exact rational execution targets.
• Adaptation: four-stage full-parameter curriculum (400 / 1320 / 3440 / 600 optimizer steps); final recovery stage observed 96,000 records.
• Final inference settings: GPT-2 tokenizer without added special tokens, sequence length 256, Euler sampling configuration with 64 recorded steps.

Training inputs and the sealed evaluation panel are intentionally not bundled with the model repository. Small reproducible smoke fixtures and integrity commitments are available in the source repository.

Evaluation

The final online weights were evaluated once on a fixed, sealed 512-problem seven-operation extrapolation panel. Only aggregate metrics are published.

Metric	Result
Final-answer accuracy	237 / 512 (46.29%)
Completion rate	444 / 512 (86.72%)
Exact oracle plan	230 / 512 (44.92%)
Syntax-valid action turns	98.24%
Previously unobserved suffix	208 / 448 (46.43%)

See metrics.json for the machine-readable aggregate report. These results measure the complete model-plus-tool protocol on this specific generated panel; they are not a general mathematics, reasoning, or natural-language benchmark.

Intended use and limitations

This release is intended for research, reproducibility, and demonstrations of discrete-diffusion action generation. Important limits include:

• It is not a general conversational assistant and may emit invalid actions.
• The demonstrated agent accepts bounded integer arithmetic expressions using +, -, *, /, and parentheses; natural-language word problems are not claimed.
• Correct calculation is performed by an external deterministic tool. The learned component plans and references calls; it does not guarantee a valid plan or final answer.
• The published score comes from one fixed synthetic extrapolation panel and should not be interpreted as broad mathematical understanding.
• Do not deploy in high-stakes settings without independent validation and appropriate safeguards.

License and attribution

The included source code retains the upstream MIT license in LICENSE. However, the upstream louaaron/sedd-medium Hugging Face repository did not state a separate model-weight license in its card metadata when this derivative release was packaged. The MIT code license must not be assumed to license the weights. Read MODEL_LICENSE.md and NOTICE.md before using or redistributing this artifact.