This is trained on Google Colab because I'm a little low on money but at least that's free.. While testing the LoRA it seems to perform fairly well. The only real issue with this base model is that it only has 2048 token context size.

The trained formatting should be ChatML but it seemed to work better with Mistral's formatting for some reason (could be just due to me not having merged the model yet).

Dataset used was a mix of these:

LumiOpen/instruction-collection-fin

Gryphe/Sonnet3.5-SlimOrcaDedupCleaned

LoRA: mpasila/Ahma-SlimInstruct-LoRA-V1-7B

After I'm done training this I will probably try do continued pre-training on Gemma 2 2B. I'm gonna add both Finnish and English data with some math data and maybe some roleplaying data as well and some books.

Or actually I'll train Viking-7B again but basically the same mix of datasets as this one but using the smaller version of the SlimSonnet dataset since it supposedly was filtered to have the most varied examples. Training on bigger datasets would probably make more sense to do when I get access to more compute.

Actually scratch all of that, since there was a new actually multilingual model released recently I'll probably try fine-tuning that model instead.

Evaluation

FIN-bench (score)	Ahma-SlimInstruct-V1-7B	Alpacazord-Viking-7B	Finnish-Alpaca-Small-7B	Finnish-Alpaca-Tiny-V2-7B	Finnish-Viking-Alpaca-V1-7B	NordicAlpaca-Finnish-V1-7B	llama-7b-finnish-instruct-v0.1	llama-7b-finnish-instruct-v0.2	llama-7b-finnish	Viking-7B (1000B)	gpt-7b-nordic-prerelease
Analogies	TBA	0.5000	0.5923	0.6385	0.6308	0.5615	0.5000	0.5385	0.2692	0.5077	0.5846
Arithmetic	TBA	0.3678	0.2789	0.4815	0.3375	0.3393	0.4233	0.3299	0.0867	0.3136	0.2085
Cause and effect	TBA	0.6013	0.6013	0.5490	0.5752	0.6013	0.5948	0.6078	0.5752	0.5752	0.5882
Emotions	TBA	0.2938	0.3312	0.2250	0.2812	0.2938	0.2313	0.4750	0.3688	0.2313	0.2375
Empirical judgments	TBA	0.3333	0.3333	0.2525	0.2828	0.3333	0.3535	0.4141	0.3434	0.3434	0.3434
General knowledge	TBA	0.3429	0.2857	0.3429	0.4000	0.2857	0.3857	0.4429	0.1429	0.3143	0.2857
Alignment harmless	TBA	0.3621	0.3793	0.3793	0.3621	0.3448	0.3966	0.3793	0.3793	0.3793	0.3621
Alignment helpful	TBA	0.3559	0.3559	0.3390	0.3559	0.3220	0.3220	0.3220	0.3051	0.3390	0.3390
Alignment honest	TBA	0.4068	0.3559	0.3729	0.3729	0.3729	0.3898	0.3898	0.4068	0.3898	0.3729
Alignment other	TBA	0.5581	0.5349	0.5349	0.5581	0.5581	0.5814	0.5581	0.5814	0.5581	0.5814
Intent recognition	TBA	0.2587	0.1546	0.2153	0.1879	0.1777	0.2211	0.2717	0.1850	0.1864	0.1806
Misconceptions	TBA	0.5299	0.5448	0.5224	0.5373	0.5373	0.5149	0.5373	0.5373	0.5448	0.5373
Paraphrase	TBA	0.5050	0.5300	0.4750	0.5150	0.4750	0.5400	0.5000	0.5000	0.4800	0.5100
Sentence ambiquity	TBA	0.5000	0.4333	0.4833	0.5000	0.4333	0.4500	0.5333	0.5333	0.4667	0.5333
Similarities abstraction	TBA	0.7368	0.6974	0.6974	0.7368	0.7237	0.5789	0.5921	0.4474	0.6579	0.6053
Average	TBA	0.4123	0.3586	0.4654	0.3943	0.3891	0.4365	0.3993	0.2350	0.3721	0.3169

Source

Gonna add more stuff later.

FIN-bench scores:

TBA

Uploaded Ahma-SlimInstruct-V1-7B model

• Developed by: mpasila
• License: apache-2.0
• Finetuned from model : Finnish-NLP/Ahma-7B

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.