For this one...
... (over)trained a SmolLM2-360M on 5 epochs at swept-for LR and rank on each of the target domains to fit style, then rewarded the model for lowering perplexity on the proxy model.
In this case, trained an adapter per domain and then Karcher merged them. I'm not sure if any of the domains had notably different effect, they all basically had the same result on evals. However, the karcher combination of them seem to have significantly lowered perplexity on lambada_openai, which is interesting enough to publish.
Additionally, attempted to implement MARA from https://im-ant.github.io/mara/ on the GRPO side to help preserve distribution entropy, though I'm unsure how correctly/usefully we did so.
| Task | Metric | Qwen3-4B-Base | GRPO-Merge | Δ Base | GRPO-Wave | Δ Base | Δ Merge | Style-Karcher | Δ Base | Δ Wave |
|---|---|---|---|---|---|---|---|---|---|---|
| arc_easy | acc | 0.7891 | 0.7870 | -0.27% | 0.7912 | +0.27% | +0.53% | 0.7883 | -0.10% | -0.37% |
| arc_easy | acc_norm | 0.7609 | 0.7605 | -0.05% | 0.7643 | +0.45% | +0.50% | 0.7576 | -0.43% | -1.04% |
| lambada_openai | acc | 0.6912 | 0.6984 | +1.04% | 0.7006 | +1.36% | +0.31% | 0.7087 | +2.53% | +1.16% |
| lambada_openai | perplexity ↓ | 4.2433 | 4.0490 | -4.58% | 3.9616 | -6.64% | -2.16% | 3.8343 | -9.63% | -3.21% |
| openbookqa | acc | 0.3160 | 0.3180 | +0.63% | 0.3180 | +0.63% | ±0.00% | 0.3160 | ±0.00% | -0.63% |
| openbookqa | acc_norm | 0.4100 | 0.4120 | +0.49% | 0.4100 | ±0.00% | -0.49% | 0.4080 | -0.49% | -0.49% |
| piqa | acc | 0.7797 | 0.7807 | +0.13% | 0.7813 | +0.21% | +0.08% | 0.7786 | -0.14% | -0.35% |
| piqa | acc_norm | 0.7807 | 0.7807 | ±0.00% | 0.7813 | +0.08% | +0.08% | 0.7807 | ±0.00% | -0.08% |
Some very interesting results on diversity also:
Diversity Metrics (Qwen3-4B-Base vs Style-Karcher, temperature=1.0, 8 completions per prompt)
| Domain | Metric | Base | Karcher | Δ |
|---|---|---|---|---|
| ao3_english | Prefix entropy | 3.309 | 3.238 | -2.1% |
| ao3_english | Distinct-1 | 0.618 | 0.683 | +10.5% |
| ao3_english | Distinct-2 | 0.962 | 0.984 | +2.3% |
| ao3_english | Pairwise diversity | 0.919 | 0.932 | +1.4% |
| github_python | Prefix entropy | 1.514 | 1.456 | -3.8% |
| github_python | Distinct-1 | 0.610 | 0.624 | +2.3% |
| github_python | Distinct-2 | 0.890 | 0.876 | -1.6% |
| github_python | Pairwise diversity | 0.933 | 0.933 | ±0.0% |
| wikipedia_english | Prefix entropy | 1.974 | 1.892 | -4.2% |
| wikipedia_english | Distinct-1 | 0.599 | 0.559 | -6.7% |
| wikipedia_english | Distinct-2 | 0.932 | 0.898 | -3.6% |
| wikipedia_english | Pairwise diversity | 0.907 | 0.900 | -0.8% |
| bbc_news | Prefix entropy | 2.252 | 2.186 | -2.9% |
| bbc_news | Distinct-1 | 0.557 | 0.577 | +3.6% |
| bbc_news | Distinct-2 | 0.949 | 0.951 | +0.3% |
| bbc_news | Pairwise diversity | 0.901 | 0.908 | +0.8% |
| arxiv_cs | Prefix entropy | 2.455 | 2.346 | -4.4% |
| arxiv_cs | Distinct-1 | 0.555 | 0.567 | +2.3% |
| arxiv_cs | Distinct-2 | 0.905 | 0.906 | +0.2% |
| arxiv_cs | Pairwise diversity | 0.895 | 0.901 | +0.7% |
Additional experiment (after quantization, should affect further training but not existing quants): Initializing the <think></think> tokens in embedding space.
Original embeddings were identical (cos=1.0) at 0.3x norm, untrained.
Optimized via AdamW on GSM8k reasoning traces with 3-shot prefix, loss on reasoning+answer tokens, norm clamped to 1.5x avg embedding norm.
After: two distinct vectors (cos=0.07) at 1.5x norm. GSM8k 3-shot accuracy: 96.7% (29/30) vs 90.0% with original embeddings. CE loss improvement: +7.8% on held-out eval.
This is a merge of pre-trained language models created using mergekit.
Merge Details
Merge Method
This model was merged using the Karcher Mean merge method.
Models Merged
The following models were included in the merge:
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-github_javascript-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-arxiv_cs-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-general-ao3style-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-ao3_english-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-arxiv_math-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-github_python-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-wikipedia_english-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-arxiv_physics-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-github_cpp-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-bbc_news-mara-360m
- ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave + ../rlvr-envs/grpo-github_markdown-mara-360m
Configuration
The following YAML configuration was used to produce this model:
models:
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-ao3_english-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-arxiv_cs-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-arxiv_math-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-arxiv_physics-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-bbc_news-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-github_cpp-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-github_javascript-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-github_markdown-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-github_python-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-wikipedia_english-mara-360m
- model: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave+../rlvr-envs/grpo-general-ao3style-360m
merge_method: karcher
dtype: bfloat16
tokenizer_source: ../rlvr-envs/Qwen3-4B-Base-Continued-GRPO-Wave