proofkit-gpt-oss-20b-lora

ProofKit gpt-oss-20b — LoRA (distillation teacher)

An attention-only LoRA adapter on openai/gpt-oss-20b, fine-tuned on the ProofKit SFT set (visproj/proofkit-sft, 2 epochs). This is the teacher in ProofKit's distillation pipeline: it generates the completions in visproj/proofkit-distill-qwen0.5b, which a Qwen 0.5B student then learns from.

It is not served in the app directly (too large for a free Space) — its quality reaches users through the distilled 0.5B GGUF.

Usage (base + adapter)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "visproj/proofkit-gpt-oss-20b-lora")
tok = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")

For high-throughput distillation the adapter is merged and served via vLLM (see scripts/modal_distill.py in the project repo).

Evaluation (post-fix, 3-judge panel)

Mean score (0–100) on 15 held-out prompts, graded by Claude Opus 4.7, GPT-5.5, and a local Qwen-3B (gpt-oss experts is a deliberately un-retrained stale control):

Both served retrained 0.5Bs beat the stale control and every untuned base across all three judges, and the distilled 0.5B ≈ ties its own 20B teacher.

About ProofKit

ProofKit is a work-sample generator for job seekers — it turns a target role, background, and skills-to-prove into a realistic, clearly-fictional practice work sample (a role-specific challenge, a guided builder, a readiness review, and a recruiter-ready portfolio packet). Built for the Hugging Face Build Small Hackathon (Backyard AI track). Integrity rules are load-bearing: outputs never claim real employment, metrics are labeled hypothetical, and exports carry an ethical disclosure.

The ProofKit model family

A note on training data (the "static responses" fix)

An earlier version of these models produced repetitive, input-ignoring drafts. The root cause was synthetic-data leakage: the dataset rendered the example user answers and the target from the same template slots, so the model learned target = template instead of target = f(input). The fix — faithfulness anchors (a distinctive token shared by the answer and the target) + seeded per-example variation across every task, then a full-chain retrain — is what these current weights reflect.

Prompt format is a frozen contract

These 0.5B models were trained on the exact prompt shapes from ProofKit's prompt_formats.py. They only behave well when prompted in that format; reworded or free-form prompts push them off-distribution. They are purpose-built components of the ProofKit app, not general chat models.