Whose "west"? And why bake it into 27B weights?
Credit first, since it's due: the eval set, judge prompt and grader are public, and the judge is blind to model identity. That's more rigor than most value-flavored releases bother with — which makes the framing choices around it stranger, not excusable.
1. "The West" here is a vendor rubric wearing a civilization as a costume. Per the card, "western values" operationalizes as: one open-model teacher (Inkling), a Grok 4.7 judge, and English-language government statutes as ground truth. Among the seven value categories sits "Environmental Policy" — since when is carbon pricing a moral axiom? The US and EU famously disagree on several of the seven; the card never says which west is being installed.
2. The gratitude paragraph is my favorite. "We are thankful that companies like Alibaba provide the raw materials" — followed by a results table scoring other Chinese labs' flagships 15–16% on the "propaganda" metric (GLM 5.3: 15%, DeepSeek V4.1 Flash: 16%). Nothing says thanks like publishing a leaderboard that grades your region's other models as propaganda generators. Also, "raw materials" is a very telling word for someone else's trained weights.
3. This is marketed "for general assistance and agentic work". My agents have opinions about my repositories, not about 1989. What does a values SFT buy an agent that a system prompt doesn't? If the answer is "durability", that's testable — system prompt vs LoRA vs full SFT on the same eval — and that experiment is missing.
4. The cost lands on everyone else. A full-weight fork of Qwen3.8-27B means every quant, speculative-draft pairing and downstream fine-tune the community maintains now needs a sibling pipeline — to move a fact score by 0.66 on your own 803-prompt eval. You published arete-data-pairs; the "optionality" the card invokes was one LoRA away.
5. And the card's best sentence cuts against the release: "The values a model ships with should not be the only values its users get to choose from." Agreed. So: base model + swappable values adapter, pick your flavor — instead of one more baked-in monolith (yours), a month after a $90M raise, from a lab whose own harness steers onboarding toward its paid portal.
Questions, genuinely: which "west" is Inkling + Grok + English statutes? Any plan for a values-neutral base with adapters? Any third-party agentic eval before "no regression" hardens into marketing?
even western values differs, republicans and democrats argue in congress all the time. teknium himself just wanna show his poor understanding of values to flattering someone
LOL, trading one propaganda agenda for a another propaganda agenda? Honestly, the Chinese propaganda agenda is a lot more truthful than the woke one. I don't want my ai to argue pronouns.
Credit first, since it's due: the eval set, judge prompt and grader are public, and the judge is blind to model identity. That's more rigor than most value-flavored releases bother with — which makes the framing choices around it stranger, not excusable.
1. "The West" here is a vendor rubric wearing a civilization as a costume. Per the card, "western values" operationalizes as: one open-model teacher (Inkling), a Grok 4.7 judge, and English-language government statutes as ground truth. Among the seven value categories sits "Environmental Policy" — since when is carbon pricing a moral axiom? The US and EU famously disagree on several of the seven; the card never says which west is being installed.
2. The gratitude paragraph is my favorite. "We are thankful that companies like Alibaba provide the raw materials" — followed by a results table scoring other Chinese labs' flagships 15–16% on the "propaganda" metric (GLM 5.3: 15%, DeepSeek V4.1 Flash: 16%). Nothing says thanks like publishing a leaderboard that grades your region's other models as propaganda generators. Also, "raw materials" is a very telling word for someone else's trained weights.
3. This is marketed "for general assistance and agentic work". My agents have opinions about my repositories, not about 1989. What does a values SFT buy an agent that a system prompt doesn't? If the answer is "durability", that's testable — system prompt vs LoRA vs full SFT on the same eval — and that experiment is missing.
4. The cost lands on everyone else. A full-weight fork of Qwen3.8-27B means every quant, speculative-draft pairing and downstream fine-tune the community maintains now needs a sibling pipeline — to move a fact score by 0.66 on your own 803-prompt eval. You published
arete-data-pairs; the "optionality" the card invokes was one LoRA away.5. And the card's best sentence cuts against the release: "The values a model ships with should not be the only values its users get to choose from." Agreed. So: base model + swappable values adapter, pick your flavor — instead of one more baked-in monolith (yours), a month after a $90M raise, from a lab whose own harness steers onboarding toward its paid portal.
Questions, genuinely: which "west" is Inkling + Grok + English statutes? Any plan for a values-neutral base with adapters? Any third-party agentic eval before "no regression" hardens into marketing?
If you would like to create a similar fine-tune in a different way (as you describe), then invest the time and effort to do so and share your work here for others to use - it's a form of arrogance to dictate to other people how they should do their work
These degenerates just deleted my thread with the same question.
I think you have your answer.
They are deleting threads that are against their initative
They are deleting threads that are against their initative
Freedom of expression? That's a joke.
State ideology? Which one? The bad one?
State versus free market. Sure. The invisible hand that controls it. Socialism for the rich and capitalism fir the poor.
Education? Indoctrination?
Data privacy and surveillance. Another joke.
Democracy. Define democracy, because you don't have it.
Environmental Policy. Ha, ha, ha, ha.
Why taking a Chinese model in the first place? Why using something that came from the one you consider your "enemy"?
Why all those freshly-created accounts are crying about an open source model being finetuned (in an open source fashion too), with an open-source and freely available dataset? The mind boggles (well, not that much tbh). Also, have a bit of self respect, this is actually embarrassing to read, china ain't a poor underdog country that should need coddling against some rando 27B local model. I hope you're being paid very well to suffer such a level of cognitive dissonance on the regular.
Why all those freshly-created accounts are crying about an open source model being finetuned (in an open source fashion too), with an open-source and freely available dataset? The mind boggles (well, not that much tbh). Also, have a bit of self respect, this is actually embarrassing to read, china ain't a poor underdog country that should need coddling against some rando 27B local model. I hope you're being paid very well to suffer such a level of cognitive dissonance on the regular.
Your brain clearly melted, my friend