closestfriend/brie-v2-7b

Brie v2 7B

Brie v2 7B is a LoRA adapter for Qwen/Qwen2.5-7B-Instruct, trained on the full 1,213-row snapshot of the researcher-authored philosophy and creative-writing corpus (the same file as the Llama adapter; the Qwen 2.5 0.5B/3B adapters used its 1,153-example subset — see the paper's Section 7.2).

Paper: Human-Curated Data Authoring with LLMs: A Small-Data Approach to Domain Adaptation.

Evaluation

The comprehensive suite contains 57 in-domain judgments over 25 unique prompts (all 57 generation pairs freshly regenerated for the 7B) and 15 separate out-of-domain controls. Every judge received blind, position-randomized A/B comparisons. The primary convention retains ties in the denominator; decisive rates are reported only as sensitivity analyses.

Panel	Primary, ties included	Decisive sensitivity
2025: GPT-4o + Gemini 2.5 Flash Lite	93.0%	95.5%
2026: Claude Haiku 4.5 + Claude Sonnet 4.5 + GPT-5 + Gemini 3.1 Pro	85.5%	86.3%

The previous higher decisive headline was a double-rounding artifact. The exact mean of the decisive fractions rounds once to 95.5%.

Across the four 2026 judges, the prompt-level counts are 39/57 unanimous Brie, 47/57 strict Brie majority, and 49/57 Brie plurality after including two 2-Brie/1-base/1-tie vote patterns. The 49/57 figure is not a strict-majority count.

The 2026 Anthropic judgments were routed through OpenRouter. Two earlier direct-Anthropic files contain only API errors and are excluded from all reported results.

Scaling boundary

The scaling result depends on the matched panel. On the shared 2025 GPT-4o and Gemini panel, ties included, Qwen 0.5B / 3B / 7B score 78.9% / 93.9% / 93.0%. On the exact shared 2026 Haiku, Sonnet, and GPT-5 subset, 7B leads 3B by 84.2% to 80.1%. The evidence supports a large 0.5B-to-3B gain and a 7B lead under the available 2026 overlap, not a universal size-performance ordering.

The n=15 out-of-domain control found performance near parity in aggregate, but is underpowered for categorical claims about retained capability. Coding is the clearest observed weakness.

Pass identities, raw counts, tie policy, and immutable hashes are documented in docs/QWEN_7B_JUDGE_PROVENANCE.md in the paper repository.

Training details

• Base model: Qwen/Qwen2.5-7B-Instruct
• Method: LoRA, rank 16, alpha 32, dropout 0.1
• Target modules: q_proj, v_proj
• Training corpus: the full 1,213-row researcher-authored corpus, including the 60 validation rows (see the paper's Section 7.2)
• Training: 1 epoch (152 steps); this published adapter is the final checkpoint-152, promoted to the repository root on 2026-06-02 and loaded as-is by the paper's 2026-06-14 evaluation — evaluated and published weights are identical
• Language: English

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-7B-Instruct"
adapter_id = "closestfriend/brie-v2-7b"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

This is a research artifact specializing one author's English-language philosophical and creative register. Verify factual claims independently.

Resources

• Paper and versions: https://doi.org/10.5281/zenodo.17657737
• Public method/code mirror: https://github.com/closestfriend/efficient-domain-adaptation
• Base model: https://huggingface.co/Qwen/Qwen2.5-7B-Instruct

Citation

@misc{karman_brie_domain_adaptation,
  author = {Karman, Hunter N.},
  title = {Human-Curated Data Authoring with LLMs: A Small-Data Approach to Domain Adaptation},
  doi = {10.5281/zenodo.17657737},
  url = {https://doi.org/10.5281/zenodo.17657737}
}