proofkit-qwen0.5b-7k

ProofKit Qwen 0.5B — fine-tuned (direct SFT)

Qwen/Qwen2.5-0.5B-Instruct fine-tuned directly on the ProofKit SFT set (visproj/proofkit-sft, ~7,000 synthetic examples). LoRA-trained, then merged to standalone weights. This is the in-Space Transformers option in ProofKit (runs on ZeroGPU / a small GPU; loads lazily on first generation).

It learns ProofKit's task contracts — section drafting, co-author drafting from rough user answers, revision actions, and strict-JSON scenario / recommendation / readiness / portfolio generation.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

m = "visproj/proofkit-qwen0.5b-7k"
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")  # carries the chat template
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.float16).to("cuda").eval()

# NOTE: prompt it with ProofKit's trained prompt_formats.py shapes — see below.
messages = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": PROMPT}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
ids = tok(text, return_tensors="pt").to("cuda")
out = model.generate(**ids, max_new_tokens=600, do_sample=True, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][ids["input_ids"].shape[-1]:], skip_special_tokens=True))

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.

Limitations

• 0.5B capacity. Reliably carries trained-style specifics, but can miss a truly arbitrary novel token and occasionally garbles. ProofKit adds a runtime fallback to a hosted instruct baseline when a draft drops the user's answers.
• Prompt-format-frozen (see below) — not a general chat model.

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.