Downloads · 30 days
0
jensjepsen/danish-lm-400m-grpo-mixed-v1
danish-lm-400m-grpo-mixed-v1 is a text generation model from jensjepsen. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
Top-3 best-of snapshots from a GRPO run that interleaved GSM8K + IFEval-DA (combined our-46 + Google-schema constraints) as a per-example dispatched reward. Base: jensjepsen/danish-lm-400m-sft-v31-avg-top3.
Downloads · 30 days
0
Access
Public
Updated Aug 14, 2026
Repo size
5 GB
Likes
0
Public
Click a slice to open those files.
.safetensors5 GB · 100%
From the Hugging Face model README
Top-3 best-of snapshots from a GRPO run that interleaved GSM8K + IFEval-DA (combined our-46 + Google-schema constraints) as a per-example dispatched reward. Base: jensjepsen/danish-lm-400m-sft-v31-avg-top3.
Three ckpts live in this repo under best1-step3375/, best2-step4375/, best3-step3250/. Composite rank score = ifeval-da prompt-strict + inst-strict + gsm8k pass@1 (each ∈ [0, 1]).
from transformers import AutoModelForCausalLM, AutoTokenizer
REPO = "jensjepsen/danish-lm-400m-grpo-mixed-v1"
SUBFOLDER = "best1-step3375" # or best2-step4375, best3-step3250
model = AutoModelForCausalLM.from_pretrained(REPO, subfolder=SUBFOLDER)
tok = AutoTokenizer.from_pretrained(REPO, subfolder=SUBFOLDER)
Chat format (single-turn): <|user|>{q}<|end|><|assistant|>, stop on <|end|>.
reward_gsm8k on gsm8k rows, reward_ifeval_combined (our 46 constraints + Google IFEval schema) on IF rows.jensjepsen/danish-if-grpo-combined-v1 (10k) interleaved with jensjepsen/danish-gsm8k:sft:train (~7.5k), 50/50 mix.jensjepsen/danish-lm-400m-grpo-if-combined-v1-step2400 (combined-run peak) into the mixed schedule; the three snapshots here are from mid-mixed training.ifeval-da p-strict + i-strict + gsm8k pass@1 (each ∈ [0, 1]). Best-K tracker in scripts/train_grpo_verifier.py (top-3 rolling, model+tokenizer only, no optimizer).Baseline column is jensjepsen/danish-lm-400m-sft-v31-avg-top3 (from its README).
danish-foundation-models/ifeval-da, n=539)Uses our patched Google verifier chain — DFM renamed change_case:capital_letters / change_case:lowercase_letters / detectable_format:constrained_response_with_argument without publishing verifier code; patches document divergence.
| metric | base | best1-3375 | best2-4375 | best3-3250 |
|---|---|---|---|---|
| composite | — | 1.028 | 1.019 | 1.014 |
| prompt-strict | 21.2 | 29.9 | 28.2 | 29.3 |
| prompt-loose | 22.0 | 30.6 | 29.9 | 29.7 |
| inst-strict | 35.2 | 45.0 | 44.9 | 44.5 |
| inst-loose | 35.8 | 45.6 | 46.1 | 44.9 |
| eval | base | best1-3375 | best2-4375 | best3-3250 |
|---|---|---|---|---|
| GSM8K pass@1 (n=1317) | 17.39 | 24.37 | 25.89 | 24.60 |
| SciQ open-Q pass@1 (n=1000) | 13.50 | 14.10 | 13.60 | 13.80 |
| CIT-gen (n=720) | 29.86 | 28.6 | 28.5 | 29.2 |
| textman_summary chrF++ | 41.11 | 40.67 | 40.84 | 40.94 |
| textman_rewrite chrF++ | 46.51 | 47.24 | 46.80 | 47.20 |
| eval | base | best1-3375 | best2-4375 | best3-3250 |
|---|---|---|---|---|
| CITMC (n=720) | 48.19 | 49.58 | 50.14 | 48.75 |
| SciQ-MC (n=1000) | — | 59.6 | 59.5 | 59.6 |
| PIQA (n=100 full test) | 53 (n=50) | 59.0 | 59.0 | 59.0 |
| ARC-Easy chat-MC | 44.40 | 42.68 | 42.55 | 42.55 |
| ARC-Challenge chat-MC | 29.35 | 29.01 | 28.92 | 28.33 |
| eval | base | best1-3375 | best2-4375 | best3-3250 |
|---|---|---|---|---|
| ARC-Easy logprob (n=2376) | 40.61 | 40.45 | 40.07 | 40.45 |
| ARC-Challenge logprob (n=1172) | 27.47 | 27.30 | 27.56 | 27.47 |
Tight — all three ~within 1pp on most metrics. Per-metric winners:
Rough guidance: pick best1 if IF adherence is primary; best2 for math; best3 for the most-balanced knowledge preservation.
Weights only (no optimizer/scheduler) — for eval/inference.