EmberForge-3B-Reasoner

Private finetuned Nanbeige4.1-3B reasoning release by strykes.

Included Artifacts

• Merged full model (Safetensors) at repo root for HF benchmarking
• LoRA adapter in adapter/
• GGUF in gguf/:
  • Nanbeige4.1-3B-Q5_K_M.gguf
  • Nanbeige4.1-3B-Q4_K_M.gguf
  • Nanbeige4.1-3B-f16.gguf
• Optional archive in archives/

Training Snapshot

• Base: Nanbeige/Nanbeige4.1-3B
• Method: Unsloth QLoRA -> merged weights
• Data: ~3.5k synthetic reasoning samples
• Epochs: 2
• Sequence length: 4096

Notes

• Intended for research and benchmarking.
• Validate outputs before critical use.

Benchmarks (2026-02-24)

Local lm-eval results (this finetune)

Task	Metric	Score
mmlu	acc,none	59.98%
gsm8k	exact_match,flexible-extract	62.40%
arc_challenge	acc_norm,none	31.74%
hellaswag	acc_norm,none	56.07%
winogrande	acc,none	50.04%
piqa	acc_norm,none	63.22%
boolq	acc,none	74.37%
truthfulqa_mc2	acc,none	45.34%

Public references

• Base model (Nanbeige/Nanbeige4.1-3B) author-published benchmarks are listed in:
  • benchmarks/lm-eval-2026-02-24/benchmark_comparison_public_2026-02-24.md
• Frontier references (Claude/GPT/Gemini) are included in the same comparison report.

Reproducibility artifacts

• benchmarks/lm-eval-2026-02-24/summary_v3.tsv
• benchmarks/lm-eval-2026-02-24/results_2026-02-24T00-06-21.474293.json
• benchmarks/lm-eval-2026-02-24/run_v3.log
• benchmarks/lm-eval-2026-02-24/benchmark_comparison_public_2026-02-24.md

Caveat

Public model-card comparisons are not always apples-to-apples with lm-evaluation-harness settings (prompting, few-shot, decoding, and benchmark versions can differ).