I just pushed the test program, you can find it here:
https://github.com/kmatheussen/radium/tree/master/faust_llm_test

It's made by an LLM to be used by an LLM, but it can probably be
useful for a human as well.
Here's an example of usage:

kjetil@Kjetils-MacBook-Pro faust_llm_test % ./faust_llm_test --help
Usage: faust_llm_test [options] <prompt> [<prompt> ...]
  --code <file>          current program (default: FaustDev2 default program)
  --effect / --mono      effect conventions (mono implies effect)
  --inputs <n>           expected effect input count (-1 = infer from prompt)
  --new-session          first prompt is creation (like "Clear history")
  --compile-error <file> seed the first request with this compiler error
  --prompt <text>        add a prompt (same as a positional argument)
  --prompt-file <file>   one prompt per line, appended to positional prompts
  --output <file>        final program (default: stdout)
  --lint-only <file>     print the textual static-analysis findings
for a program and exit
  --lint-compile <file>  print the compile-based static-analysis
findings for a program and exit
  --compile-only <file>  compile a program and print its channel counts and exit
  --program-dir <dir>    Radium program dir (default <exe>/../bin)
  --log <file>           LLM log path (default $RADIUM_LLM_LOG or
~/.radium/llm.log)
  --quiet

kjetil@Kjetils-MacBook-Pro faust_llm_test % ./faust_llm_test
--lint-compile /tmp/chowning_fix_9.faust
Line 20: the definition 'chowningReverb' does not compile on its own,
so it is part of the compile error.
Line 20: the definition 'chowningReverb' is part of a self-referential
cycle - Faust has no assignment, so 'x = ... x ...;' is an endless
evaluation cycle, not feedback. Build the loop with the '~' operator:
'loop = + ~ (de.delay(maxSamples, delaySamples) : *(feedback));
process = input : loop;'. For a modulated feedback delay:
'modDelay(sig) = de.delay(2 * ma.SR, (0.03 + 0.005 * os.osc(0.5)) *
ma.SR, sig); process = _,_ : par(i, 2, + ~ (modDelay : *(0.7)));'.
Line 44: the definition 'mod1' does not compile on its own, so it is
part of the compile error.
Line 44: the definition 'mod1' uses 'base1', which is never defined
anywhere in the program - this gives the 'undefined symbol' error. It
is probably a leftover from an example or a previous revision: either
define it (e.g. base1 = hslider("base1", 0.5, 0, 1, 0.01);) or remove
every use of it.
Line 45: the definition 'mod2' does not compile on its own, so it is
part of the compile error.
Line 45: the definition 'mod2' uses 'base2', which is never defined
anywhere in the program - this gives the 'undefined symbol' error. It
is probably a leftover from an example or a previous revision: either
define it (e.g. base2 = hslider("base2", 0.5, 0, 1, 0.01);) or remove
every use of it.

kjetil@Kjetils-MacBook-Pro faust_llm_test % ./faust_llm_test
--lint-only /tmp/chowning_fix_9.faust
LLM: Indexed 1340 Faust library definitions (1700 symbols, 864 in
compact table, 19404 chars) from
/Users/kjetil/radium/bin/packages/faust/libraries
Line 28: 'delay1' is defined inside a 'with' block but never used -
dead code. Remove it, or use it in the function body (e.g. return it)
- as written, the function computes something else.
Line 29: 'delay2' is defined inside a 'with' block but never used -
dead code. Remove it, or use it in the function body (e.g. return it)
- as written, the function computes something else.

On Tue, Sep 15, 2026 at 2:34 PM Kjetil Matheussen
<[email protected]> wrote:
>
> Yes, that's very interesting. I have to try faust-rs to see if it
> reduces number of retries in Radium. Note that Radium already tries to
> give better error messages back to the LLM when compilation fails,
> that's one of the main things it has been trained for, but I have
> difficulty believing having even better error messages from Faust in
> the first place would increase the number of retries it has to make.
>
> On Mon, Sep 14, 2026 at 5:56 PM Stéphane Letz <[email protected]> wrote:
> >
> > Interesting. You may be interested by incant-audio project: 
> > https://github.com/Losera/incant-audio
> >
> > and in particular this discussion: 
> > https://github.com/Losera/incant-audio/issues/26
> >
> > It would be interesting to see if « better » error messages for the Faust 
> > compiler can help in your use-case.
> >
> > Stéphane
> >
> > > Le 14 sept. 2026 à 15:18, Kjetil Matheussen <[email protected]> a 
> > > écrit :
> > >
> > > I have a command line program that sends LLM requests the same way
> > > Radium does. A much stronger LLM runs this command line program over
> > > and over,
> > > both for prompts I ask it to do, but usually prompts it generates itself.
> > > For each round it analyzes the returned program, compiles it, and
> > > based on that improves the information sent to the LLM when asking it
> > > to generate Faust code. It also
> > > trains code that analyse what went wrong if compilation fails, or
> > > there is something else that looks wrong with the code, and this
> > > information is sent back to
> > > the LLM to generate Faust code when something goes wrong. These files
> > > are auto-trained:
> > > https://github.com/kmatheussen/radium/blob/master/bin/faust_library_guide.txt
> > > https://github.com/kmatheussen/radium/blob/master/audio/Faust2_llm_lint.cpp
> > > https://github.com/kmatheussen/radium/blob/master/audio/Faust2_llm_lint.hpp
> > > https://github.com/kmatheussen/radium/blob/master/Qt/LLM_client.hpp
> > > https://github.com/kmatheussen/radium/blob/master/Qt/LLM_faust_session.hpp
> > >
> > >
> > > On Mon, Sep 14, 2026 at 8:40 AM Stéphane Letz <[email protected]> wrote:
> > >>
> > >> Hi Kjetil,
> > >>
> > >> Thanks for the infos. It would be great if you could explain how you do 
> > >> «  auto-training » to have the models work without thinking.
> > >>
> > >> Thanks.
> > >>
> > >> Stéphane
> > >>
> > >>> Le 13 sept. 2026 à 21:51, Kjetil Matheussen <[email protected]> 
> > >>> a écrit :
> > >>>
> > >>> Radium 8.1.49 has been released now.
> > >>>
> > >>> This is a big update for Faust:
> > >>> * FaustDev2: Lots of auto-training fixes, trained both on
> > >>>  deepseek flash (no thinking) and chat-gpt 5.6 luna (no
> > >>>  thinking). This means that the LLM prompt almost always
> > >>>  creates valid and working programs even for larger
> > >>>  tasks such as "autotune" when using cheap and
> > >>>  non-reasoning/thinking models.
> > >>> * Faust: Fix for exponentially increasing compilation time
> > >>>  and svg generation time for various faust code.
> > >>> * Faust: Add patch to fix endless recursion in some programs.
> > >>> * Faust: Add patch to fix hang for programs which has
> > >>>  "a function argument is a shared block-diagram DAG".
> > >>> * Faust: Upgrade 2.81.2 -> 2.88.0
> > >>>
> > >>> _______________________________________________
> > >>> Faudiostream-users mailing list
> > >>> [email protected]
> > >>> https://lists.sourceforge.net/lists/listinfo/faudiostream-users
> > >>
> >


_______________________________________________
Faudiostream-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/faudiostream-users

Reply via email to