Closed pabl-o-ce closed 1 year ago
The Gradio app for this one at https://huggingface.co/spaces/openaccess-ai-collective/rlhf-arena is GGML based and super slow, takes 30+ seconds to get a completion result.
Looks like it's USER/ASSISTANT style with a sorta weird system prompt: Below is a dialogue between a USER and an ASSISTANT. The USER may ask questions, request information, or provide instructions for a task, often supplementing with additional context. The ASSISTANT responds accurately and effectively, offering insights, answering questions, or executing tasks to the best of its ability based on the given information.
Going to give TheBloke's GPTQ and GGML quants a go as this model does look very interesting.
Looks like this model comes in 3 rather confusing flavors:
Minotaur-15B is based on StarChat Minotaur-13B is based on LLaMa (but had a problem with dataset) Minotaur-13B-fixed is also based on LLaMa
I have previously evaluated Minotaur-13B
but that doesn't seem to even be valid anymore since that model has been superseded by a fix.
The GPTQ version seems to be missing a file required to load the tokenizer - opened a PR https://huggingface.co/TheBloke/minotaur-15B-GPTQ/discussions/1
Missing file has been added and model now works with the safer use_fast=False
tokenizer
Some interesting answers out of this guy, it seems to have been excessively instruction-tuned:
Here's one way you can write this JavaScript function `substrCount`:
1. First, create a variable called `result` which will store the count of occurrences of the substring `substr`. Initialize it as zero (0).
2. Next, use the built-in method `indexOf()` to check if the string `str` contains the substring `substr`. If so, increment the value stored in `result` by 1.
3. Finally, return the final value of `result`, indicating how many times the substring occurred within the input string.
This will score a 0 on our test, since its not code.
15b eval has been added but I'm going to leave this issue open for now as I'd like to also try minotaur-13b-fixed.
I have been testing TheBloke/minotaur-15B-GGML
with this prompt you have prompts/Minotaur.txt
using koboldcpp
and is very solid answers using the prompt template
and the config params
that you have to generate questions.. I don't know if that helps on the testing procedure that you have.
That's very interesting because that prompt is meant for a totally different model, Minotaur-13B, which is based on Llama not StarCoder 🧐
I'll give it a run though and see what happens..
On Sun, Jun 25, 2023, 1:30 PM PΔBLØ ᄃΞ @.***> wrote:
I have been testing TheBloke/minotaur-15B-GGML with this prompt you have prompts/Minotaur.txt using koboldcpp and is very solid answers using the prompt template and the config params that you have to generate questions.. I don't know if that helps on the testing procedure that you have.
— Reply to this email directly, view it on GitHub https://github.com/the-crypt-keeper/can-ai-code/issues/21#issuecomment-1606173682, or unsubscribe https://github.com/notifications/unsubscribe-auth/AUGCACCSEJZR3HS2NLJKJ7DXNBYUFANCNFSM6AAAAAAZSV4WDE . You are receiving this because you commented.Message ID: @.***>
Llm are interesting! indeed.
Should I create new issues for Orca
and the new Wizard uncensored
created from faldore
Also I really like to test code questions... I experience whenever I have a bad answer.. I try to regenerate and sometimes give me some correct answer.. just to keep in mind as a fact I wanna share.
I did some work with Orca Mini family this weekend, it's sitting on a branch and I need to rerun the evals to be JS only but here's a fun little experiment: https://huggingface.co/spaces/mike-ravkine/orca-mini-coder-analysis
Falcoder is on my wishlist but nobody has made quants and I can't easily run it full precision.
On Sun, Jun 25, 2023, 11:51 PM PΔBLØ ᄃΞ @.***> wrote:
Llm are interesting! indeed.
Should I create new issues for Orca and the new Wizard uncensored created from faldore
Also I really like to test code questions... I experience whenever I have a bad answer.. I try to regenerate and sometimes give me some correct answer.. just to keep in mind as a fact I wanna share.
— Reply to this email directly, view it on GitHub https://github.com/the-crypt-keeper/can-ai-code/issues/21#issuecomment-1606549767, or unsubscribe https://github.com/notifications/unsubscribe-auth/AUGCACB6CZFOXU2AQRJ4COTXNEBNJANCNFSM6AAAAAAZSV4WDE . You are receiving this because you commented.Message ID: @.***>
Falcoder is on my wishlist but nobody has made quants and I can't easily run it full precision.
Let me send the repo mrm8488/falcoder-7b
to TheBloke if he can made it...
UPDATE: he is going to do it.
I am impatient, so I've done it myself - q5_0 uploading now: https://huggingface.co/mike-ravkine/falcoder-7b-GGML
I've tested with this GGML fork and it seems to work: https://github.com/jploski/ggml/tree/falcon40b
$ ./bin/falcon -m ~/ai/falcoder-7b-GGML/ggml-model-falcoder-7b-sharded-bf16-q5_0.bin --top_k 1 -p 'function fib('
main: seed = 1687870044
falcon_model_load: loading model from '/home/miner/ai/falcoder-7b-GGML/ggml-model-falcoder-7b-sharded-bf16-q5_0.bin' - please wait ...
falcon_model_load: n_vocab = 65024
falcon_model_load: n_embd = 4544
falcon_model_load: n_head = 71
falcon_model_load: n_head_kv = 1
falcon_model_load: n_layer = 32
falcon_model_load: ftype = 2008
falcon_model_load: qntvr = 2
falcon_model_load: ggml ctx size = 4764.96 MB
falcon_model_load: memory_size = 32.00 MB, n_mem = 65536
falcon_model_load: ........................ done
falcon_model_load: model size = 4732.91 MB / num tensors = 196
extract_tests_from_file : No test file found.
test_gpt_tokenizer : 0 tests failed out of 0 tests.
main: number of tokens in prompt = 3
main: token[0] = 5529, function
main: token[1] = 10045, fib
main: token[2] = 19, (
function fib(n) {
if (n === 0) {
return 0
}
if (n === 1) {
return 1
}
return fib(n-1) + fib(n-2)
}
console.log(fib(5)) // 8
console.log(fib(6)) // 13
console.log(fib(7)) // 21
console.log(fib(8)) // 28
console.log(fib(9)) // 34
console.log(fib(10)) // 55
console.log(fib(11)) // 78
console.log(fib(12)) // 121
console.log(fib(13)) // 169
console.log(fib(14)) // 233
main: mem per token = 424356 bytes
main: load time = 1802.23 ms
main: sample time = 97.24 ms
main: predict time = 47944.04 ms / 237.35 ms per token
main: total time = 50607.29 ms
A few initial thoughts:
main: number of tokens in prompt = 4
main: token[0] = 602, int
main: token[1] = 4278, []
main: token[2] = 10045, fib
main: token[3] = 19, (
int[] fib(int n) {
int a = 0, b = 1, c = 0;
for (int i = 1; i < n; i++) {
c = a + b;
a = b;
b = c;
}
return [a, b];
}
int main() {
int n = 5;
int[] fib = fib(n);
for (int i = 0; i < fib.length; i++) {
printf("%d ", fib[i]);
}
printf("\n");
return 0;
}
### Solution:
int main() {
int n = 5;
int[] fib = fib(n);
for (int i =
OK so following up on this, Minotaur-15B has been added (as per this original ticket)
Opened up #24 to track falcoder, its got some issues at first glance.
You might also be interested in #23 which also has unique challenges.
Hi guys love the work.
I have been testing
TheBloke/minotaur-15B-GGML
ans is pretty solid you can test TheBloke/minotaur-15B-GPTQ