06-reference

alphasignal nous research qwen3 debias tencent terminal rsi

2026-11-09·reference·source: AlphaSignal·by AlphaSignal editorial team
open-weight-modelsmodel-biaschina-aiai-safetyharness-engineering

Why this is in the vault

Nous Research published a reproducible, open-weights case study showing political bias in Qwen3 can be removed with plain supervised fine-tuning alone — no RL, no full retrain — which is a concrete data point for how fragile/shallow alignment-layer behavior is in open-weight Chinese models, framed in the same issue against Tencent's self-retraining Terminal-RSI model as two opposite responses to the same 27B weight class.

Mapping against Ray Data Co

Most concrete connection: this is exactly the kind of reproducible bias-audit framework that would back a client conversation in phData's enterprise-AI advisory work when a client is evaluating a Chinese open-weight model (Qwen, DeepSeek, GLM) for production — Nous Research's result (Qwen3.8-27B → Arete-27B via SFT on 53M tokens, propaganda rate 22%→<1%, fact score 2.32→2.98/5, general benchmark performance flat at 65.60% vs 65.77% on Hermes Bench) is a citable, falsifiable example that de-biasing doesn't require RL or a full retrain, and that general capability doesn't have to degrade to fix it. That's a sharper, numbers-backed version of the "alignment/bias sits in a shallow, swappable layer" claim this vault has been accumulating evidence for across the open-weight-models thread (06-reference/2026-07-19-alphasignal-inkling-pragmatic-open-weights-engineering.md, synthesis-harness-thesis-dissent-2026-04-12.md). Secondary item (Tencent's Terminal-RSI, a 27B model retraining itself from its own wins with zero human input) is adjacent-interesting for the standing Jev/self-improving-model due-diligence thread but doesn't cross an active-project threshold today — no RDCO workstream currently depends on self-retraining model internals.

Curation section

Zero deep-fetches triggered — every item was adequately covered by AlphaSignal's own blurb text; nothing cleared the "third-party + RDCO-relevant + specific unanswered hook" bar for a follow.

⚠️ Sponsorship

Three paid placements, all previously confirmed members of AlphaSignal's rotating sponsor pool (per 01-projects/process-newsletter/README.md, pool now 39+ confirmed entrants) — no new entrants this issue:

Masthead "In Partnership with" slot present but unresolved — plaintext renders only "In Partnership with ·" immediately before "Today's Author," with no legible name or link, consistent with the majority-unresolved pattern already logged in the README (most issues since 2026-09-14 render a generic/blank masthead; only 2026-09-09 and 2026-09-25 resolved cleanly). Not guessed here.

Bias implication: standard AlphaSignal rotating-pool pattern — treat sponsor blocks and adjacent editorial framing as commercially motivated by default; no new-relationship flag needed this issue.

Related