[HN Gopher] GPT-OSS Reinforcement Learning
       ___________________________________________________________________
        
       GPT-OSS Reinforcement Learning
        
       Author : vinhnx
       Score  : 147 points
       Date   : 2025-09-27 02:01 UTC (20 hours ago)
        
 (HTM) web link (docs.unsloth.ai)
 (TXT) w3m dump (docs.unsloth.ai)
        
       | Der_Einzige wrote:
       | I love love love Unsloth and everything they do, so do not take
       | what I am about to say as criticism of them.
       | 
       | But what's the point? GPT-OSS is regarded as a pretty bad open
       | source model compared to the latest deepseek or qwen releases.
       | Most attempts to use Reinforcement Learning or even any kind of
       | post-training fail in that the data you have is of worse quality
       | and quantity than the data that the model was originally trained
       | on. So you get catastrophic forgetting and a model with lower
       | general IQ than before fine-tuning.
       | 
       | This is true btw even if you use lora or better techniques to
       | supposedly "mitigate" catastrophic forgetting. Even pyreft/reft,
       | which in some cases impact only 0.001% of a models parameters,
       | cause these kind of issues in my experiments.
       | 
       | So why should anyone except AI researchers and the big 4 AI
       | providers care about fine-tuning? The vast majority of people who
       | think they need fine-tuning need good quality RAG/Agentic RAG
       | systems, since they can trivially add or remove data to their
       | model (machine unlearning doesn't work yet), also ground models
       | and objectively makes them more accurate, and fully manipulate
       | and manage how it's used in their prompts context. On top of
       | that, vector DBs/embeddings "easily" scale to billions of
       | records.
        
         | vlovich123 wrote:
         | I have compared instruction following of stock gpt-OSS against
         | stock qwen and the 20B outperformed all others, intelligently
         | following instructions and reasoning about tool calling
         | correctly for tools it hasn't been trained on. Additionally, it
         | performs like a 3B model because it uses 32 experts. I don't
         | know where this claim that it sucks comes from but my
         | evaluation of similarly competitive models showed it leading
         | the pack by a lot.
        
           | Der_Einzige wrote:
           | Even if you're right about all of this it still doesn't
           | refute my "fine-tuning doesn't work for 99% of customers
           | including you" thesis.
           | 
           | Also, most mainstream AI benchmarks do not agree with you:
           | 
           | LLMarena (https://lmarena.ai/leaderboard) has GPT-OSS 120B as
           | #53 and GPT-OSS 20B as #69 (nice), which is extremely far
           | from leading.
           | 
           | DeepSeek V3.1 is ranked #9, and is a solid 60+ elo points
           | above GPT-OSS.
           | 
           | I know you're going to link some of the "ya but chatbot arena
           | sucks cus of theoretical attacks against it" paper and the
           | llama4 debacle, but here's more evidence that GPT-OSS blows:
           | 
           | https://livebench.ai/#/?q=GPT-oss
           | 
           | GPT-oss global average: 54.60
           | 
           | Deepseek V3.1 thinking global average: 70.75
           | 
           | Qwen 3 32B global average: 63.71
           | 
           | So bring receipts next time because I did.
        
             | vlovich123 wrote:
             | Qwen 32B is a dense model. The competitors for GPT-OSS 20B
             | are max 8B dense models. You're comparing it against models
             | it's not competing with and calling it crap. That's like
             | saying Ferrari is better than Toyota. Sure, but only if
             | we're comparing 0-60. If we add a budget of 20K, suddenly
             | the Toyota starts looking more competitive.
             | 
             | I never claimed it was a frontier model. Just best in class
             | for the performance it can achieve and the memory footprint
             | it can fit in.
             | 
             | And btw OSS did super well on domain specific tests without
             | fine tuning. A model I don't need to fine tune beats one
             | that does.
        
               | Der_Einzige wrote:
               | You don't mention that MoE models horrible damage
               | logprobs (by definition) and basically need a whole new
               | theory of LLM sampling written for them.
               | 
               | Dense models are better for a reason, and the idea that
               | "everyone is doing MoE now and dense models are dead" is
               | total bunk nonsense.
               | 
               | You can quantize dense models, and 4 bit quantized Qwen
               | 32B is still better than full precision GPT-OSS. Luckily
               | Unsloth even gives you tools to go down to 1.58bits!
        
               | stingraycharles wrote:
               | Excuse my ignorance, but what is a "dense" model? Why is
               | comparing GPT-OSS 20B against 8B dense models "fair" but
               | not when comparing it against 32B models?
        
               | jychang wrote:
               | So gpt-oss-20b is a sparse MoE model
               | 
               | Which means it has ~3b parameters active per token.
               | 
               | Qwen3-32b has 32b params active per token
        
               | spwa4 wrote:
               | An MoE model will not activate most of the model for any
               | given query, so there is zero compute happening across an
               | increasing part of the model.
               | 
               | Dense model means 32B parameters => 32B get used in
               | calculation for every token. Every calculation takes
               | time, and assuming similar latent space size (which they
               | all have). For example Qwen-32B
               | 
               | MoE model has for example 80B parameters, but only 3B get
               | used in calculation for any given token. For example
               | Qwen3-Next 80B A3B
               | 
               | Performance comparison:
               | 
               | Qwen-32B => 56 tok/sec, 32 GB of VRAM
               | 
               | Qwen3-Next 80B A3B => 167 tok/sec, 85 GB of VRAM
               | 
               | So despite being close to 3x "bigger", Qwen3-Next is more
               | than 3 times faster with the same compute capacity.
               | There's a but though. But because what gets activated
               | from one token to the next is a different subset of the
               | model, it is still critical to have all 80B parameters
               | loaded into memory.
               | 
               | So MoE performs much better with less compute, at the
               | cost of more memory. It also performs better on
               | benchmarks, it is a better model.
               | 
               | Similar techniques have long been used in ML to great
               | success, rather than trying to create one brilliant
               | model, create many that each have pros and cons, and then
               | train a second model to figure out the best model for the
               | task in front of you. There's even a name for the
               | practice "ensemble models". MoE only kind-of fits
               | (because you can't easily swap out models)
               | 
               | There's other factors, the big one being attention.
               | That's why non-attention models, like MAMBA, will wipe
               | the floor in terms of performance per flop (a compute
               | unit), with anything else. When it comes to intelligence
               | however ...
        
               | yowlingcat wrote:
               | Qwen3 Next and Qwen3-30b-a3b are pretty decent proxies
               | for GPT-OSS 120B and 20B respectively (and in fact are
               | both MoEs with 3B active rather than 8B active
               | parameters), and they lap GPT-OSS pretty hard in this
               | specific benchmark, getting to #17 and #33 respectively.
               | That being said, it's hard to take benchmarks beyond a
               | grain of salt because real world tasks that I try to use
               | these models for always have a lot more variation than
               | the benchmarks illustrate. I do view GPT-OSS as a pretty
               | good alternative to the Qwen models in some cases but
               | there are tradeoffs -- while I see better reasoning from
               | GPT-OSS sometimes, the prompt adherence and overall
               | flexibility of the Qwen models makes them a lot better
               | IMO as general purpose local open weight models.
        
             | mountainriver wrote:
             | I literally talked to 5 customers last week that needed
             | fine tuning, legitimately needed it. I get if you're just
             | doing basic RAG on text you generally don't but that's only
             | part of the ecosystem
        
           | asabla wrote:
           | I'm always so confused by those statements as well. Because
           | just like you, I feel that the 20B version is really good at
           | following instructions.
           | 
           | Some of the qwen models are too, but they seem to need a bit
           | more handholding.
           | 
           | This is of course just anecdotal from my end. And I've been
           | slacking on keeping up with evals while testing at home
        
           | corlinp wrote:
           | The performance is great, but the censorship is ridiculous
           | for me. I tried it as a backend for my game Guessix[1], but
           | it would refuse for ridiculous reasons like "Cannot answer
           | questions about copyrighted works like Harry Potter."
           | 
           | 1. https://guessix.com/
        
             | BoorishBears wrote:
             | Use constrained generation
        
               | artdigital wrote:
               | Mind explaining?
        
               | 7thpower wrote:
               | Curious as well
        
               | BoorishBears wrote:
               | If you constrain the model to a JSON schema, most
               | frivolous refusals go away.
               | 
               | And if you finetune on a few formatted examples the
               | effect is even greater
        
               | corlinp wrote:
               | Do you mean like structured outputs? Unfortunately here
               | the model is guided to explicitly tell you when you
               | violate the rules and why, it can confuse it's system
               | rules with the game rules and say you're not allowed to
               | ask a question about copyrighted material etc.
        
             | artdigital wrote:
             | Try the uncensored/jailbroken variants like openai-gpt-
             | oss-20b-abliterated-uncensored-neo-imatrix
             | 
             | I just tried to ask it how to make crystal meth and it
             | generated a very detailed step by step guide
        
               | Squarex wrote:
               | I have heard that uncensorted gpt-oss is not very good
               | because of it being trained mainly on synthetic data. Is
               | not not true?
        
               | bavell wrote:
               | Iirc abliteration (ablation?) can be done without
               | "training" and is pretty quick. It finds the individual
               | weights related to the concept you want to ablate, and
               | modifies those weights to "deactivate" them. Precision
               | brain surgery, to anthropomorphize.
        
               | Squarex wrote:
               | The problem with synthetic data would be that the
               | censored information would not be in the training data at
               | all.
        
               | corlinp wrote:
               | Very interesting! Do the benchmarks hold up well or does
               | it reduce performance in other areas too?
        
         | stingraycharles wrote:
         | Lots of business decisions are based on "we prefer to use this
         | model because it's from OpenAI", not necessarily because it's
         | the best.
         | 
         | I've also seen a lot of enterprises suddenly invest in fine-
         | tuning their own models, even though there is absolutely no
         | reason they should be doing that. But because there's a huge
         | "we need to do something with AI" directive from the top. Eg
         | "if we fine tune an AI model to understand our custom query
         | language it will prevent our data scientists from taking down
         | production databases". This is an actual example I encountered
         | just last week.
         | 
         | So if there's a point, it's probably not that it's the best
         | idea, but rather than enterprises are willing to buy it.
        
         | danielhanchen wrote:
         | Oh hey! Thanks for the love :)
         | 
         | The primary goal of the release and our notebook
         | https://colab.research.google.com/github/unslothai/notebooks...
         | was actually to showcase how to mitigate reward hacking in
         | reinforcement learning - for example when RL learns to cheat
         | and output global variables instead like editing the timer to
         | cheat on benchmarking and others. You can edit the notebook to
         | do rl on other powerful models like Qwen, Llama etc
         | automatically with Unsloth as well via our automatic compiler!
         | We also made sink attention and moe inference super optimized
         | for training - note flash attention 3 doesn't have sink
         | backwards support so you'll have to use unsloth.
         | 
         | Gpt-oss tbh in our tests is a truly powerful model, especially
         | the 120b variant - it's extremely popular in western
         | enterprises since yes it's from openai but also because
         | reasoning mode high and the censored nature and its reasoning
         | capabilities are attractive. A big underutilized feature is its
         | web search and internal intermediate tool calling which it can
         | do as part of its reasoning chain just like o3 or gpt5.
         | 
         | RL yes isn't an all powerful hammer, but it can solve so many
         | more new problems. For a financial institution, you can make
         | automatic trading strategies via RL. For an intelligence
         | agency, decryption via RL. For a legal startup, possibly case
         | breakthroughs via RL, automatic drug candidates etc. And yes,
         | big labs want to automate all tasks via massive RL for eg being
         | able to play pokemon and all other games as one example. RL
         | opened so many doors since you don't need any data, just one
         | prompt like "make fast matrix multiplications kernels", and
         | reward functions - it can allow many more interesting use cases
         | where data is a constraint!
        
           | ripped_britches wrote:
           | Can you elaborate on "decryption via RL"
        
             | terataiijo wrote:
             | I think you can train a model to decrypt an encrypted. My
             | friend tried this only on like simple example tho. As long
             | as we have the environment, we can do these things.
        
             | danielhanchen wrote:
             | Definitely not breaking any modern day standards, but from
             | what I understand, some folks are trying it on simple
             | ciphers or combinations of simple ciphers to first see if
             | RL can help.
        
           | Der_Einzige wrote:
           | I'm sorry but I don't buy for a second that you can do
           | meaningful and even close to reliable decryption with RLHF on
           | currently known secure ciphers.
           | 
           | Furthermore, I'm very worried that whoever may be paying for
           | this is barking up the wrong tree. I feel that the damage
           | done with extremely bad decryption attempts would massively
           | outweigh the very few times when whatever it "decrypts" is
           | meaningfully close to what the actual text was.
           | 
           | I'm aware of how easy certain things in surveillance are (I.e
           | n-gram analysis is enough to dox anyone on HN in like 10
           | words of text) - but even sort of decent decryption of
           | SHA-256 would be a literally front page of the world
           | achievement.
        
             | vlovich123 wrote:
             | If you're going to be rude and arrogant, then the level of
             | knowledge you exhibit has to match. SHA-256 decryption
             | would be a front of the world achievement because it would
             | be redefining foundational mathematics since it's not an
             | encryption algorithm. The words you'd be looking for are
             | either a collision of SHA-256 or breaking encryption
             | algorithms like AES, RSA, ECC etc.
             | 
             | sha-256 is used in the construction of certain encryption
             | algorithms as a primitive but by itself never encrypts
             | anything. If it did you've also got middleout compression
             | invented since you could encrypt arbitrary length input
             | into 256 bits of output.
        
               | danielhanchen wrote:
               | Oh yes if RL breaks SHA-256 that'll be revolutionary -
               | but definitely not that - some folks are for now
               | investigating basic combinations of old school ciphers
               | for now - security applications with RL are most likely
               | for now related to automatically finding attack surfaces
               | and creating defensive layers - I probably should have
               | re-worded "decryption for RL" to just "security for RL"
               | sorry!
        
         | KronisLV wrote:
         | > So why should anyone except AI researchers and the big 4 AI
         | providers care about fine-tuning?
         | 
         | I currently have Qwen3 Coder 30B A3B on prem for developers to
         | use and it's pretty good for that: fits within two Nvidia L4
         | cards (with quantization), has about 60 tok/s performance and
         | can even use tools with something like RooCode or OpenWebUI.
         | 
         | However, if anyone asks it stuff in Latvian, it does mess up
         | more often than not, like someone who has half-learned the
         | language: more or less uses the words that represent the
         | correct concepts, but not always the best pick and really often
         | missing or wrongly used diacritics (a, c, e, g, i, k, l, n, s,
         | u, z). In a word, some basis of the language is there, but
         | sadly in practice it's still unusable.
         | 
         | So far working with Latvian text I need to maintain EuroLLM
         | running in parallel, which has great Latvian knowledge, but at
         | the same time just knows less (not a model that's good for
         | programming) and doesn't know how to call tools as well and
         | isn't really meant for long contexts:
         | https://huggingface.co/collections/utter-project/eurollm-66b...
         | 
         | So my ideal model (for this use case) would be something along
         | the lines of: around 30B since can't fit anything much bigger
         | into the VRAM for now, MoE, supports tool calling, okay
         | programming knowledge, has a long enough context support, can
         | converse in Latvian so I don't need to run 2 models (which
         | means that the context sizes that can fit within VRAM are way
         | too small for either of the models).
         | 
         | Without finetuning, I just have to sit and wait for someone to
         | release that and it feels unlikely that it'll just pop into
         | existence (even the bigger EuroLLM model is nowhere to be seen,
         | TildeOpen works bad). With finetuning, maybe I have a chance to
         | take the Latvian Wikipedia or whatever other samples of the
         | language being used I can get, filter down to topics I care
         | about, maybe use EuroLLM to generate question/answer pairs for
         | training from that input data and then just run Unsloth over a
         | weekend or something (who knows, maybe that'd be enough to
         | bring Qwen3 up to speed).
         | 
         | Do most people care about that stuff? Probably not, but when
         | you have to deal with a language that's less represented in the
         | training data but will _never_ have the budget to train
         | something from scratch, finetuning is nice to have. RAG can 't
         | fix it not knowing how to use language.
        
         | mountainriver wrote:
         | This is a deeply untrue statement echoed by a lot of people
         | unfortunately.
         | 
         | There was a post the other day on HN where a dev spent $2k
         | doing RL on an open model and beat the frontier models on Web
         | Voyager.
         | 
         | Research is packed with examples like this yet we keep hearing
         | from the community that training doesn't work.
         | 
         | Then consider voice/vision modalities where context engineering
         | doesn't work well at all. I just spent 2 years on the computer
         | use problem and can tell you with absolute certainty that
         | context won't get you there (we tried and tried). The only
         | thing that meaningfully moved the needle was RL.
         | 
         | I really wish this idea that only big labs can train models
         | would die. It's really hurting the ecosystem at this point.
        
         | strangescript wrote:
         | GPT-OSS models are amazing, and a lot of the bad press was poor
         | implementations of them in the usual tools and people not
         | understanding how to handle their unique quant out of the box
         | approach. Unlsoth has done amazing job unpacking best
         | approaches
        
         | ericfr11 wrote:
         | I work for a D2C product, with a standard frontend/backend, and
         | a RAG system. It's becoming annoying when the product manager
         | (not a tech guy) keeps asking for a fine-tuned LLM: most people
         | read news and assume they know what's best.
        
       | mountainriver wrote:
       | Thank you unsloth for the amazing work once again!
       | 
       | The new sleep mode in vLLM is really amazing, and it seems the
       | community hasn't quite wrapped their heads around how much more
       | accessible this makes RL training.
       | 
       | I'm reading a lot of dismaying posts on this thread, pushing the
       | idea that only big labs should be doing RL. This couldn't be
       | further from the truth, folks should try it for themselves and
       | see the outcomes!
        
         | danielhanchen wrote:
         | Thank you! No worries at all! Yes! Sleep mode is super cool
         | since this means the allocation of memory for inference can be
         | totally decoupled away from training, which opens the door to
         | many larger RL runs!
        
       | wey-gu wrote:
       | Unsloth/Denial is amazing.
        
         | homarp wrote:
         | Daniel I suppose?
        
           | danielhanchen wrote:
           | Ye probs mis-spelt :)
        
       | decodebytes wrote:
       | You can now generate reasoning with Tool calling or format pre-
       | existing datasets to the OpenAI harmony format used for gpt-oss
       | with DeepFabric:
       | 
       | https://github.com/lukehinds/deepfabric/discussions/334
        
       | WOTERMEON wrote:
       | Reads like ai slop / mad marketer
        
       ___________________________________________________________________
       (page generated 2025-09-27 23:01 UTC)