The second-generation local workshop release, built on Qwen 3.5 9B and tuned for stronger everyday reasoning, stricter JSON/tool behavior, and solid false-premise correction. V2 is trained from fresh CUDA-native adapters that do not reuse earlier MLX adapters.

What This Release Is

Full fused PyTorch weights from the best V2 checkpoint
Base: Qwen/Qwen3.5-9B
Workshop lineage: v55 targeted SFT
CUDA-native PEFT training (no MLX adapter reuse)

Comparisons

Internal workshop evals — directional, not leaderboard claims.

Eval	Stock Qwen3.5-9B	V1 (`Qwenjamin_Franklin`)	V2 (this)
`full40`	309/400	325/400	341/400
`false_smoke`	102/110	110/110	97/110
`json_hard`	15/30	30/30	30/30
`tool_schema_canary`	50/175	106/175	169/175
`workbench_local_agent`	63/100	72/100	70/100
`no_tool_leakage`	99/100	100/100	99/100

What improved vs V1: full40 reasoning (+16 pts), tool schema discipline (+63 pts). What V1 still leads on: false-premise correction (110 vs 97), workbench (72 vs 70).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("stamsam/Qwenjamin_Franklin_V2", torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("stamsam/Qwenjamin_Franklin_V2", trust_remote_code=True)

Notes

For strict tool JSON, use explicit output instructions.
Verify important outputs in high-stakes workflows.
A compact 4-bit MLX version is at stamsam/Qwenjamin_Franklin_V2_4bit.