Kinshield Bonsai d24-w3200 — v1.0.0-r08-20260924
Experimental NAS/pruning/full LoRA recovery artifact. Created using Bonsai by Prism ML.
Derived from original SFT, revision 011ee6c8d05d49144fa5747bedf23e8904883564.
Contents and size
q1/base-Q1_0.gguf + q1/adapter-f16.gguf: 169895008 bytes combined. Cap applies to this deployment pair, not this whole repository.model/: matching pruned HF base and tokenizer, necessary for this adapter. Do not attach the adapter to the unpruned original base.hf_adapter/: final R08 LoRA weights; metadata/: pruning map, recipe, certificates and provenance.contract/: original trained contract including recommendations. benchmark/: OOD comparison, recommendations ignored only in core-v1 scoring.
Recovery
One full epoch, 77,358 original train samples, 2,418 updates, effective batch32, 3,283,235 assistant tokens. LoRA recovery from R07 warm adapter; frozen child base. Not full-parameter finetuning. Train/dev/OOD separation preserved. Checkpoint, base unchanged, cold replay and Q1 certification PASS.
Verify and use
Download a pinned release tag or commit, then run python scripts/verify.py.
Q1 requires the pinned PrismML llama.cpp fork/build in metadata/engine-materialization.json; generic llama.cpp compatibility is not asserted. Set LLAMA_SERVER to the matching server and run bash scripts/serve.sh. Supply contract/system_prompt.txt as system message and the expected SMS/evidence JSON as user message; greedy decoding, thinking disabled, context16384, max1024 output tokens.
HF loading (requires compatible Transformers/PEFT):
from pathlib import Path
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
root = Path('/path/to/downloaded/release')
base = AutoModelForCausalLM.from_pretrained(root / 'model', local_files_only=True)
model = PeftModel.from_pretrained(base, root / 'hf_adapter', local_files_only=True)
tokenizer = AutoTokenizer.from_pretrained(root / 'model', local_files_only=True)
Explicitly load the packaged base: adapter_config's original relative base path alone is not a portable auto-loader instruction.
Evaluation and limitations
See paired OOD core results. Each model: entity384 and scenario384 samples. Strict JSON/verdict/finding catalog validation; recommendations excluded symmetrically. Same raw generation under old contract, not a model retrained for a new two-field contract. D26 leads Findings F1; D24 leads observed verdict F1. SCAM support only36/24 per split. This release validates the pipeline; selection-dev/final winner audit remains incomplete. No raw training/evaluation examples or optimizer checkpoints are uploaded.
Original certificates record source paths as provenance; use metadata/checksums.json for portable package verification. Do not overwrite this release; changes require a new version/tag.
