v2 vs v1
Hi. I accidentally found this keyword model and was impressed with the result.Good job! She will definitely help me with my studies when the external Internet is turned off. Tell me, what is the difference between versions 1 and 2 in terms of the model's "understanding" of the theory.
The interpretation, generally, should be about the same (it was trained on the same data from the same model with the same understanding of doctrine), however it should be more general because more generalized data was included; furthermore it will attempt to be more rhetorically convincing in its answers, it will better be able to use statistics to prove its arguments, and it will be able to more adequately combat the foundational myths capital builds as opposed to working under its constraints. At least that was the goal, I'd make no guarantees it fully worked, but imo this model is much better than the original so :p
Okay, I'll compare it! Do you have any plans to train on similar data to qwen3.5 ? Thank you so much for your work!
Maybe in the future, I'm not too sure on Qwen 3.5 as a finetuning base yet
I've been testing it extensively in recent days, and I have something to say about it. As far as basic knowledge of dialectics and the classics goes, everything's fine. But then the model starts to stumble, when applied to the history of the USSR after the 1920s. It drifts heavily into Trotskyism, and the logic breaks down altogether. It comes to the wonderful conclusion that the USSR managed to close its technological gap thanks to slave labor, lol. It also denies the need for industrialization and the development of heavy industry at that time. Furthermore, it uncritically cites CIA documents and the works of modern bourgeois historians as sources, calling them "independent" without sufficient evidence, along with a whole bunch of other deconstructed anti-Soviet myths from the 1980s and 1990s, in the spirit of Solzhenitsyn. Based on this, I think the information for training needs to be critically re-examined. I'm using a translator, so it's a bit awkward in places, but I hope I've conveyed the general idea.
Interesting!
How did you test? Most of my training data doesn't involve proper exchanges with the model, only single turn questions into answers, so I definitely wouldn't be shocked if it falls apart in that context
Furthermore, it uncritically cites CIA documents and the works of modern bourgeois historians as sources, calling them "independent" without sufficient evidence
I think I know where this comes from and it can probably be solved by just being more careful with what data I tell it it can use to be honest. I have some better ideas for how to improve the data quality and add proper dialectic reasoning to the model so I would quite like to come back to this when I get a chance
It also denies the need for industrialization and the development of heavy industry at that time.
welcome back mao /j
Like I've been focusing on more general purpose models recently but I have more knowledge to come back to this and make a good domain model for theory
Well, I tested it by asking questions manually. Of course, lol. I'm studying Marxism myself, and I was asking about things I already understand well enough to spot errors in reasoning. And in context, the model was already responding to specific questions where the aforementioned jokes regularly surfaced. Moreover, the basic logic defined by the classics is generally preserved, but since the model can't apply them to self-critique the facts it knows, it ends up with contradictions it doesn't see and defends. I also noticed that it sometimes completely breaks down on everyday questions, which isn't a problem for a specialized model, but the answers are extremely funny. Try asking it riddles, you'll die laughing. If it could be further trained with corrected information, something like qwen 3.5-9B would be a fantastic assistant for beginners studying Marxism, one that could explain non-obvious points of the theory.
I almost forgot: perhaps it would make sense to use "Reasoning" models as the foundation? If we view this LLM as a modern-day analogue to the *Manifest der Kommunistischen Partei *, then it certainly wouldn't hurt to try and squeeze every last drop of capability out of the limited hardware on which it will be running. I think there is some sense to this.
Speaking of Mao, reading his works remains a rather distant prospect for me at the moment; however, I did come across a critical analysis of his activities authored by Čchen Šao-jü —and the opening section, in particular, proved quite telling. It was rife with a staggering number of dishonest polemical tactics, logical fallacies, and similar manipulative devices aimed at the unsophisticated reader. Moreover, in contemporary China, Mao does not command any particular authority—and this fact, combined with the country’s current capitalist trajectory, allows for certain preliminary conclusions to be drawn. All in all, the reports from the top Chinese leader are quite curious: business, business, business—and then, somewhere right at the end, a couple of words about the proletariat.