Included trained weights
adapter_model.safetensors: ordinary LoRA, r128 / alpha256.
head_tokens.safetensors: 17 trained start/end/null token rows; E-series also restores per-layer embedding rows.
tokenizer/: trained tokenizer and native Gemma template.
grouped_expert_lora.safetensors: additionally required for 26B-A4B (expert r16 / alpha32).
modeling.py, ADAPTER_FORMAT.json: the actual custom reconstruction logic.
PEFT loading alone is incomplete. Use the provided loader to restore special rows, PLE and grouped expert deltas where present. Files are the completed training artifacts, without optimizer/RNG or copied base weights. Training completed 19041 steps, one epoch; data fraction is a corpus selection fraction, not pruning.
Recorded quality
Frozen bilingual SimpleTool Diagnostic-800, 400 Chinese + 400 English, four 200-case slices and 400 related pairs. Historical local typed literal matcher; not official BFCL AST/execution. Greedy, max128 new tokens per head, native trained turn template, BF16 base and autocast with FP32 adapters/rows, SDPA+KV. These are recorded measurements from 2026-11-17, not a fresh Hub reload benchmark.
Table with columns: Metric, Result| Metric | Result |
|---|
| Status | Diagnostic only; no ability score |
No completed 800-case ability score is reported. Repeated invalid function/argument output was reproduced during diagnosis; this checkpoint is released for inspection, not recommended deployment.
The required-call matcher also requires schema validity; malformed optional values can fail it. Independent parameter branches can misbind entities or correlated slots. Related sample pairs are not IID tasks. Vision and content quality were not evaluated.
Inference
Use the original environment's model support: Python3.12, Torch2.11.0 CUDA12.8, transformers5.17.0 and PEFT0.21.3, plus safetensors/accelerate. Newer versions require independent validation. Obtain the original base locally and download this whole repository.
python infer.py --base /absolute/path/to/original-base \ --artifact /absolute/path/to/this-repository --messages messages.json --args 2
messages.json is a JSON list of system/user messages containing answer-free tool definitions and the query. The native chat template closes thought for the larger family as trained. --content opts into content generation. This reference decodes each branch sequentially for clarity and does not reproduce the CUDA parallel speed path. Output is a proposal and no tools execute.
Full code/evaluation: https://github.com/HaxxorCialtion/SimpleTool . Training corpus publication is being prepared separately; the diagnostic dataset is not the training corpus. Base/model/code licensing follows the accompanying Apache2.0 license and upstream notices.