Downloads · 30 days
7
5% of all-time downloads
jensjepsen/danish-lm-400m-sft-v31-avg-top7
danish-lm-400m-sft-v31-avg-top7 is a machine learning model from jensjepsen. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
400M-parameter Danish SFT model (v31, top-7 ckpt weight-space average). The 7 constituent checkpoints are also present as subfolders (see below).
Downloads · 30 days
7
5% of all-time downloads
All-time downloads
154
Public
Parameters
415M
13.3 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors13.3 GB · 100%
From the Hugging Face model README
400M-parameter Danish SFT model (v31, top-7 ckpt weight-space average). The 7 constituent checkpoints are also present as subfolders (see below).
jensjepsen/danish-lm-400m-base-ropext8048-v1 (400M, 8048-token context, RoPE-extended)danish-arc:sft:train + danish-openbookqa:sft:train)DataCollatorWithFlattening)adamw_bnb_8bit, LR 3e-5 constant + 500 warmupEach constituent is at <step-N>/ in this repo (no optimizer state, model + tokenizer only).
Chat-format, generation-based. n = full test split.
| eval | avg-top7 | avg-top3 | best single (step 42732) |
|---|---|---|---|
| gsm8k (0-shot pass@1) | 17.08 | 17.39 | 15.69 |
| sciq | 14.40 | 13.50 | 12.53 |
| citgen | 27.64 | 29.86 | 27.36 |
| citmc | 48.19 | 48.19 | 47.51 |
| mean-4 | 0.268 | 0.272 | 0.258 |
| eval | avg-top7 | avg-top3 |
|---|---|---|
ARC-Easy chat-MC (jensjepsen/danish-arc:arc_easy:test) | 44.32 | 44.40 |
ARC-Challenge chat-MC (jensjepsen/danish-arc:arc_challenge:test) | 30.20 | 29.35 |
| ARC-mean chat-MC | 37.26 | 36.88 |
OpenBookQA chat-MC (jensjepsen/danish-openbookqa:main:test) | 36.00 | 35.40 |
| PIQA chat-MC (2-choice, random 50) | 51.00 | 53.00 |
| GPQA-Diamond chat-MC (4-choice, random 25) | 22.73 | 21.72 |
| textman_summary chrF++ | 42.06 | 41.11 |
| textman_rewrite chrF++ | 47.13 | 46.51 |
| eval | avg-top7 | avg-top3 |
|---|---|---|
| SciQ logprob (n=1000) | 45.40 | 44.90 |
| CITMC logprob (n=720) | 59.58 | 58.61 |
| ARC-Easy logprob (n=2376) | 41.46 | 40.61 |
| ARC-Challenge logprob (n=1172) | 27.05 | 27.47 |
avg-top7 wins on 3/4 logprob metrics — broader averaging pool helps the raw discrimination signal.
Multi-turn few-shot (prefixing 5 <|user|>Q<|end|><|assistant|>A<|end|> pairs) breaks this model — GSM8K drops from ~18 → 2%. Model was trained single-turn; the fake past turns confuse it. Use 0-shot or in-context examples inside a single user turn.
danish-foundation-models/ifeval-da)Format-compliance eval — SWA averaging tends to slightly hurt format-heavy metrics.
| metric | avg-top7 | avg-top3 | v31 step-24927 (mid-run) | v22-avg (prior best) |
|---|---|---|---|---|
| prompt-strict | 22.0 | 21.2 | 24.1 | 22.7 |
| prompt-loose | 22.9 | 22.0 | 25.4 | – |
| inst-strict | 34.8 | 35.2 | 39.4 | 36.6 |
| inst-loose | 35.5 | 35.8 | 40.6 | – |
danish-metamath-gsm:sft, danish-algebra-sft-v5-mixed, danish-arith-chain-sft-v1, danish-wiki-grounded-sft-v3:sft, danish-text-to-question-v2:sft, danish-sciq:sft:train, danish-gsm8k:sft:train, danish-instruction-following-v4:sft:train, danish-wiki-closedqa-v1:sft, danish-word-problems-v2, danish-wiki-closedqa-stem-v1:sft, danish-wiki-broadqa-stem-v1:sft, danish-wiki-mc-letters-v1, danish-rc-v1, danish-reason-v1, danish-textman-v1, danish-arc:sft:train, danish-openbookqa:sft:train
<|user|>{prompt}<|end|><|assistant|>{answer}<|end|>