Downloads · 30 days
0
jensjepsen/danish-lm-400m-grpo-mixed3-intermediary
danish-lm-400m-grpo-mixed3-intermediary 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.
Three top best-of snapshots from a GRPO run that interleaves GSM8K + IFEval-DA (our-46 + Google-schema) + JSON schema tasks as a per-example dispatched reward. Base: jensjepsen/danish-lm-400m-grpo-json-intermediary-st…
Downloads · 30 days
0
Access
Public
Updated Aug 15, 2026
Repo size
5 GB
Likes
0
Public
Click a slice to open those files.
.safetensors5 GB · 100%
From the Hugging Face model README
Three top best-of snapshots from a GRPO run that interleaves GSM8K + IFEval-DA (our-46 + Google-schema) + JSON schema tasks as a per-example dispatched reward. Base: jensjepsen/danish-lm-400m-grpo-json-intermediary-step3125 (already saturated on JSON; this run rebuilt IF+gsm8k while preserving JSON).
Three ckpts live under best1-step625/, best2-step500/, best3-step875/. Composite rank = sum of ifeval_da_prompt_strict + ifeval_da_inst_strict + gsm8k_pass@1 + json_mean_reward (all from the greedy-eval callback during training).
Intermediary — this is a mid-run snapshot (625/39260 steps ≈ 0.8%). Full offline eval numbers are for best1-step625 only; best2/best3 reported only via the in-training composite.
from transformers import AutoModelForCausalLM, AutoTokenizer
REPO = "jensjepsen/danish-lm-400m-grpo-mixed3-intermediary"
SUBFOLDER = "best1-step625" # or best2-step500, best3-step875
model = AutoModelForCausalLM.from_pretrained(REPO, subfolder=SUBFOLDER)
tok = AutoTokenizer.from_pretrained(REPO, subfolder=SUBFOLDER)
Chat: <|user|>{q}<|end|><|assistant|>, stop on <|end|>.
Note: gsm8k below is first-200 subset, systematically easier than the full test. Real gsm8k number is in the full-suite table further down.
| metric | best1-625 | best2-500 | best3-875 |
|---|---|---|---|
| composite (sum) | 2.152 | 2.148 | 2.130 |
| ifeval-da prompt-strict (541 rows) | 30.1 | 29.5 | 30.2 |
| ifeval-da inst-strict (541 rows) | 46.3 | 45.6 | 45.9 |
| gsm8k pass@1 (first-200) | 30.5 | 31.5 | 29.0 |
| json mean_reward (200 held-out) | 108.4 | 108.2 | 107.8 |
best1-step625vs baseline danish-lm-400m-sft-v31-avg-top3 (SFT with no GRPO) and prior danish-lm-400m-grpo-mixed-v1-best1-step3375 (GSM8K + IF only, no JSON).
danish-foundation-models/ifeval-da, n=539)| metric | v31 base | grpo-mixed-v1 | mixed3-best1 |
|---|---|---|---|
| prompt-strict | 21.2 | 29.9 | 28.9 |
| prompt-loose | 22.0 | 30.6 | 30.1 |
| inst-strict | 35.2 | 45.0 | 45.9 |
| inst-loose | 35.8 | 45.6 | 46.6 |
| eval | v31 base | grpo-mixed-v1 | mixed3-best1 |
|---|---|---|---|
| GSM8K pass@1 (full n=1317) | 17.39 | 24.37 | 25.89 |
| SciQ open-Q pass@1 (n=1000) | 13.50 | 14.10 | 14.80 |
| CIT-gen (n=720) | 29.86 | 28.6 | 28.3 |
| textman_summary chrF++ | 41.11 | 40.67 | 40.69 |
| textman_rewrite chrF++ | 46.51 | 47.24 | 45.30 |
| eval | v31 base | grpo-mixed-v1 | mixed3-best1 |
|---|---|---|---|
| CITMC (n=720) | 48.19 | 49.58 | 48.1 |
| SciQ-MC (n=1000) | — | 59.6 | 59.5 |
| PIQA (n=100) | 53 (n=50) | 59.0 | 57.0 |
| ARC-Easy (n=2376) | 44.40 | 42.68 | 41.92 |
| ARC-Challenge (n=1172) | 29.35 | 29.01 | 28.41 |
| GPQA-Diamond (n=198) | — | 21.21 | 23.74 |
| eval | v31 base | grpo-mixed-v1 | mixed3-best1 |
|---|---|---|---|
| ARC-Easy logprob (n=2376) | 40.61 | 40.45 | 40.57 |
| ARC-Challenge logprob (n=1172) | 27.47 | 27.30 | 26.62 |
| GPQA-Diamond logprob (n=198) | — | 25.76 | 26.77 |
jensjepsen/ifbench-da-v1, n=300)Verified on the AI2 IFBench extension of Google's IFEval registry — 58 additional verifier classes covering counts, palindromes, permutations, csv formats, format-line-indent, etc.
| metric | grpo-mixed-v1 | mixed3-best1 |
|---|---|---|
| prompt-strict | 10.3 | 10.3 |
| inst-strict | 11.3 | 10.8 |
| aggregate (avg of prompt-strict + prompt-loose) | 11.83 | 11.67 |
jensjepsen/danish-json-grpo-v1:eval, n=200)Entirely new capability vs prior mixed-v1 (which had zero JSON training). Reward is graded parse + key-set + gold-value match:
| metric | grpo-mixed-v1 | mixed3-best1 |
|---|---|---|
| mean reward (max 1.2 for gold-having rows) | — | 1.084 (94% of 1.15 weighted ceiling) |
esperanto_lm.rl_rewards.reward_mixed dispatches per-row — reward_gsm8k on gsm rows, reward_ifeval_combined on IF rows, reward_json_schema on json rows (parse + key-set match + gold-value match + null penalty).jensjepsen/danish-if-grpo-combined-v1 (10k IF) + jensjepsen/danish-gsm8k:sft:train (~7.5k) + jensjepsen/danish-json-grpo-v1 (~9.7k). Interleaved, shuffled. Total: 27,170 rows/epoch × 3 epochs.jensjepsen/danish-lm-400m-grpo-json-intermediary-step3125 — JSON at ~94% of ceiling but IF/gsm8k drifted from pure-JSON training. The mix restored those in ~625 steps.BestCkptSaverCallback (weights only, no optimizer).Weights only (no optimizer/scheduler) — for eval/inference.