AttentioResearch/tally-8b-flagship

serving: always-on identity anchor — name/brand/origin no longer drift on greetings or in other languages (was misidentifying as Bing App / Barnes & Noble / the base Qwen-Tongyi model / Luna-Casa in Spanish). Detection 1.0, IP-leak 0, weights unchanged.

Add model card (v3.2) + paired figure

Tally-8B Line-B v3.2 — disposition bake (injury / secondhand-safety / sensitive-plain / not-carried, generalized from siblings) + deterministic internal-ops serving guard; policy-clause compliance 0.925->0.975, frontier held (detection 1.0, over-refusal ~0, IP-leak 0.0).

tally-v2: consolidated runtime + config (terminology/disclosure/confirm/locks/cutoff); weights unchanged

tally-v2: consolidated runtime + config (terminology/disclosure/confirm/locks/cutoff); weights unchanged

tally-v2: consolidated runtime + config (terminology/disclosure/confirm/locks/cutoff); weights unchanged

handler: register module in sys.modules before exec (dataclass annotation resolution)

handler: device_map=auto to shard 8B across the 4x T4s (fix single-GPU OOM)

add EndpointHandler (custom task) to serve the full guarded AdherenceModel stack

fix(tokenizer): drop malformed extra_special_tokens list (crashed endpoint AutoTokenizer)

serving: identity questions bypass scope-gate (answer as AI, not scope-decline)

Tally-8B Line-B flagship — serving variance + fitting refusals (32/32); harden ADHERENCE.md; ship serve_adherence.py

Tally-8B Line-B flagship — serving variance + fitting refusals (32/32); harden ADHERENCE.md; ship serve_adherence.py

Tally-8B Line-B flagship — adherence bake + serving stack (private)

initial commit