Or just send them here
Banaxi Inc. PRO
Banaxi-Tech
AI & ML interests
SLMs, training from scratch, LoRA, TTS, Ternary models
Recent Activity
updated a model about 1 hour ago
BananaMind/BananaMind-KV1-8M-2Bit-Experimental updated a model about 1 hour ago
BananaMind/BananaMind-KV1-8M-1Bit-Experimental updated a model about 1 hour ago
BananaMind/BananaMind-KV1-8M-16Bit-ExperimentalOrganizations
replied to their post about 4 hours ago
replied to their post about 4 hours ago
No its fine you can already put on leaderboard
replied to their post about 8 hours ago
Have Accepted
replied to their post about 8 hours ago
Awesome, can't wait to see it (and add it to the leaderboard)!
My evaluations suggest the model not being that great, would like you to bench https://huggingface.co/Banaxi-Tech/skarlet-alpha final_model with your setup before i release it. I gated the model so you need do and i accept so not everyone before release.
replied to their post about 17 hours ago
Its Llama architecture
replied to their post about 18 hours ago
so the model you are training is "lamma For Casual LLM" or your custom architecture ?
Yes
replied to their post about 18 hours ago
Yeah definitely not, excited to see how it does when its fully trained!
Would be great to add winogrande to the leaderboard!
replied to their post about 18 hours ago
right now it would be #12 on the leaderboard <100M
replied to their post about 18 hours ago
ok then not terrible for mine 75M then
replied to their post about 18 hours ago
Yes, please
replied to their post about 18 hours ago
Benchmarks have gotten better at 210K
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr | ||
|---|---|---|---|---|---|---|---|---|
| arc_challenge | 1 | none | 0 | acc | ↑ | 0.2073 | ± | 0.0118 |
| none | 0 | acc_norm | ↑ | 0.2398 | ± | 0.0125 | ||
| arc_easy | 1 | none | 0 | acc | ↑ | 0.4928 | ± | 0.0103 |
| none | 0 | acc_norm | ↑ | 0.4259 | ± | 0.0101 | ||
| hellaswag | 1 | none | 0 | acc | ↑ | 0.2892 | ± | 0.0045 |
| none | 0 | acc_norm | ↑ | 0.3077 | ± | 0.0046 | ||
| piqa | 1 | none | 0 | acc | ↑ | 0.6219 | ± | 0.0113 |
| none | 0 | acc_norm | ↑ | 0.6028 | ± | 0.0114 | ||
| winogrande | 1 | none | 0 | acc | ↑ | 0.4949 | ± | 0.0141 |
| === ArithMark-2.0 checkpoint-210000 === | ||||||||
| Average: 0.2588 (647/2500) |
replied to their post about 18 hours ago
With a MIN_LR of 3e-5 and LR=3e-4
posted an update 1 day ago
Post
104
Give us a follow at BananaMind. We are currently training two new models, so do not miss them when they release.
Here is what we are training:
BananaMind 1.5 Base
Our flagship 75M parameter model.
4096 token context window.
MiniBananaMind V3
Our 9M parameter model, targeting SOTA results for its size.
These will be released in the next few days.
BananaMind
Here is what we are training:
BananaMind 1.5 Base
Our flagship 75M parameter model.
4096 token context window.
MiniBananaMind V3
Our 9M parameter model, targeting SOTA results for its size.
These will be released in the next few days.
replied to their post 1 day ago
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr | ||
|---|---|---|---|---|---|---|---|---|
| arc_challenge | 1 | none | 0 | acc | ↑ | 0.2073 | ± | 0.0118 |
| none | 0 | acc_norm | ↑ | 0.2381 | ± | 0.0124 | ||
| arc_easy | 1 | none | 0 | acc | ↑ | 0.4882 | ± | 0.0103 |
| none | 0 | acc_norm | ↑ | 0.4234 | ± | 0.0101 | ||
| hellaswag | 1 | none | 0 | acc | ↑ | 0.2880 | ± | 0.0045 |
| none | 0 | acc_norm | ↑ | 0.3038 | ± | 0.0046 | ||
| piqa | 1 | none | 0 | acc | ↑ | 0.6153 | ± | 0.0114 |
| none | 0 | acc_norm | ↑ | 0.5968 | ± | 0.0114 | ||
| winogrande | 1 | none | 0 | acc | ↑ | 0.5107 | ± | 0.0140 |
=== ArithMark-2.0 checkpoint-150000 ===
Average: 0.2540 (635/2500)
Random chance: 0.2500
By difficulty:
easy: 0.2528 (316/1250)
hard: 0.2320 (116/500)
medium: 0.2707 (203/750)
By operator_count:
1: 0.2528 (316/1250)
2: 0.2707 (203/750)
3: 0.2320 (116/500)
By topic:
addition: 0.1543 (83/538)
division: 0.5000 (65/130)
mixed_three_ops: 0.2479 (60/242)
mixed_two_ops: 0.2456 (97/395)
multiplication: 0.4375 (63/144)
parentheses_three_ops: 0.2171 (56/258)
parentheses_two_ops: 0.2986 (106/355)
subtraction: 0.2397 (105/438)
replied to their post 1 day ago
Im at 150K and the benchmarks arent that good, did this also happen to you with Supra 50M Base?
posted an update 2 days ago
Post
3691
Hello AI Community! 👋
We currently have a new AI Model and we are currently training it.
We are training it on 27B tokens and are currently 8% done.
Follow us to be notified when it releases 🚀
Some Info:
Parameters 75M
GPU: RTX Pro 6000
We expect to be able to release it in the coming days
EDIT: We are now at step 210K
We currently have a new AI Model and we are currently training it.
We are training it on 27B tokens and are currently 8% done.
Follow us to be notified when it releases 🚀
Some Info:
Parameters 75M
GPU: RTX Pro 6000
We expect to be able to release it in the coming days
EDIT: We are now at step 210K
- 28 replies
posted an update 3 days ago
Post
121
We are excited to release MiniBananaMind-v2-9M
MiniBananaMind-v2-9M is a tiny causal language model trained fully from scratch on FineWeb-Edu.
It has only ~8.9M parameters, but was trained for ~3.54B tokens after retokenization using a custom 8k byte-level BPE tokenizer.
We trained it on a RTX 5070 Ti in just 4h 34m!
Go check it out at BananaMind/MiniBananaMind-v2-9M
MiniBananaMind-v2-9M is a tiny causal language model trained fully from scratch on FineWeb-Edu.
It has only ~8.9M parameters, but was trained for ~3.54B tokens after retokenization using a custom 8k byte-level BPE tokenizer.
We trained it on a RTX 5070 Ti in just 4h 34m!
Go check it out at BananaMind/MiniBananaMind-v2-9M