Downloads · 30 days
0
jensjepsen/danish-lm-400m-sft-v39-mid
danish-lm-400m-sft-v39-mid 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. The card lists the license as apache-2.0.
Checkpoints from Danish SFT run v39 (400M, 3 epochs, 45,336 steps). Trained on a 22-source Danish mix; the tool data is jensjepsen/danish-tool-dialogues-v5, which removes reasoning prose before tool calls and answers…
Downloads · 30 days
0
Access
Public
Updated Sep 6, 2026
Repo size
16.6 GB
Likes
0
Public
Click a slice to open those files.
.safetensors11.6 GB · 70%
From the Hugging Face model README
Checkpoints from Danish SFT run v39 (400M, 3 epochs, 45,336 steps). Trained on a
22-source Danish mix; the tool data is
jensjepsen/danish-tool-dialogues-v5,
which removes reasoning prose before tool calls and answers every dangling call.
| checkpoint | gsm8k | citgen | sciq | ifeval | icl | extraction | tool_seen | tool_unseen | tool_answer | agg |
|---|---|---|---|---|---|---|---|---|---|---|
| step-3777 | 9.6 | 19.0 | 7.0 | 34.1 | 40.1 | 19.8 | 76.9 | 69.7 | 67.4 | 0.382 |
| step-7554 | 11.2 | 21.9 | 8.1 | 35.1 | 46.6 | 24.4 | 78.2 | 67.4 | 68.5 | 0.402 |
| step-11331 | 14.0 | 19.4 | 9.1 | 33.1 | 49.1 | 24.1 | 71.5 | 66.7 | 69.9 | 0.397 |
| step-15108 | 14.9 | 20.0 | 9.1 | 37.2 | 51.3 | 22.8 | 84.6 | 74.2 | 70.4 | 0.427 |
| step-18885 | 15.1 | 21.1 | 10.0 | 35.2 | 50.0 | 21.7 | 85.1 | 75.2 | 71.5 | 0.428 |
| step-22662 | 14.7 | 22.1 | 11.5 | 34.9 | 50.8 | 24.0 | 75.4 | 65.6 | 70.6 | 0.411 |
| step-26439 | 14.1 | 23.5 | 10.6 | 36.1 | 55.7 | 26.4 | 86.3 | 80.0 | 72.9 | 0.451 |
| step-30216 | 14.8 | 22.4 | 10.7 | 35.6 | 54.0 | 26.9 | 89.7 | 80.1 | 72.0 | 0.451 |
| step-33993 | 14.2 | 20.4 | 11.1 | 35.2 | 52.0 | 30.3 | 85.6 | 74.2 | 71.7 | 0.439 |
| step-37770 | 13.4 | 24.7 | 11.8 | 37.7 | 53.3 | 29.0 | 87.0 | 73.0 | 71.7 | 0.446 |
| step-41547 | 16.4 | 22.5 | 13.5 | 37.4 | 54.0 | 29.9 | 88.1 | 71.0 | 71.6 | 0.449 |
| step-45324 | 14.4 | 21.8 | 9.1 | 34.3 | 50.3 | 28.6 | 90.4 | 80.2 | 71.7 | 0.445 |
| avg-top3 | 17.7* | 25.3* | 13.8* | 36.7* | 58.3* | 29.0* | 91.0 | 80.6 | 72.3 | — |
avg-top3 = uniform average of step-26439, step-30216, step-41547.
* measured at batch 256; every other number in the table at batch 32. Tool
metrics for avg-top3 were measured at batch 32 and are directly comparable.
tool_seen / tool_unseen are argument pair-F1 (wrong tool scores 0), scored
against danish-tool-dialogues-v5 eval splits. tool_unseen uses tools held
out by name hash. tool_answer is grounding of the reply in the tool result;
gold replies score 83.1 on this metric, a reply lifted from another row 3.4.
Prior run v38 (same mix, tool-dialogues-v4) best: tool_seen 76.8, tool_unseen
65.2, agg 0.437. v38's tool_answer is not comparable — different eval split.
| folder | files | resumable |
|---|---|---|
step-* | 11 | yes (optimizer, scheduler, RNG) |
avg-top3 | 6 | no |
final | 7 | no (step 45,336, never eval'd) |
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(
"jensjepsen/danish-lm-400m-sft-v39-mid", subfolder="avg-top3")
t = AutoTokenizer.from_pretrained(
"jensjepsen/danish-lm-400m-sft-v39-mid", subfolder="avg-top3")
Chat format: <|user|> ... <|end|> <|assistant|> ... <|end|>, with
<|tool_call|>{"name": ..., "arguments": {...}}<|/tool_call|> and
<|tool_result|>{...}<|/tool_result|>. Tools are offered as a JSON list after
Værktøjer: in the first user turn.