91 lines
4.5 KiB
Markdown
91 lines
4.5 KiB
Markdown
---
|
||
license: apache-2.0
|
||
base_model: Qwen/Qwen2.5-0.5B-Instruct
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
language:
|
||
- en
|
||
tags:
|
||
- proofkit
|
||
- distilled
|
||
- sft
|
||
- build-small-hackathon
|
||
- work-sample
|
||
datasets:
|
||
- visproj/proofkit-distill-qwen0.5b
|
||
---
|
||
|
||
# ProofKit Qwen 0.5B — distilled (merged)
|
||
|
||
`Qwen/Qwen2.5-0.5B-Instruct` **distilled** from the ProofKit gpt-oss-20b teacher
|
||
([`visproj/proofkit-gpt-oss-20b-lora`](https://huggingface.co/visproj/proofkit-gpt-oss-20b-lora)).
|
||
Sequence-level (data) distillation: the teacher's completions over ProofKit's prompts
|
||
([`visproj/proofkit-distill-qwen0.5b`](https://huggingface.co/datasets/visproj/proofkit-distill-qwen0.5b))
|
||
are used to SFT the student (LoRA, 3 epochs), then merged to standalone weights.
|
||
|
||
The **GGUF build of this model**
|
||
([`visproj/proofkit-distilled-qwen0.5b-gguf`](https://huggingface.co/visproj/proofkit-distilled-qwen0.5b-gguf))
|
||
is what the ProofKit Space serves through llama.cpp — free on CPU. This merged
|
||
Transformers copy is the source for that conversion and for evaluation.
|
||
|
||
## 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):
|
||
|
||
| model | Claude | GPT-5.5 | Qwen-3B | **Avg** |
|
||
|---|---:|---:|---:|---:|
|
||
| gpt-5.5 (frontier ceiling) | 94.6 | 95.6 | 90.8 | **93.7** |
|
||
| gpt-oss attn (retrained teacher) | 82.0 | 66.8 | 81.4 | **76.7** |
|
||
| **qwen-0.5b distilled (served)** | 79.0 | 68.6 | 82.2 | **76.6** |
|
||
| **qwen-0.5b direct 7k (served)** | 78.6 | 64.4 | 82.0 | **75.0** |
|
||
| gpt-oss experts *(stale control)* | 67.6 | 68.6 | 81.8 | **72.7** |
|
||
| qwen-3b base | 62.1 | 67.1 | 80.5 | **69.9** |
|
||
| gpt-oss base | 55.4 | 53.8 | 68.2 | **59.1** |
|
||
| qwen-0.5b base | 36.5 | 44.5 | 67.9 | **49.7** |
|
||
|
||
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; prompt-format-frozen (see below). A purpose-built ProofKit component.
|
||
|
||
## About ProofKit
|
||
|
||
[ProofKit](https://huggingface.co/spaces/visproj/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
|
||
|
||
| Repo | What it is |
|
||
|---|---|
|
||
| [`visproj/proofkit-qwen0.5b-7k`](https://huggingface.co/visproj/proofkit-qwen0.5b-7k) | Qwen2.5-0.5B fine-tuned directly on the 7k set (Transformers) |
|
||
| [`visproj/proofkit-gpt-oss-20b-lora`](https://huggingface.co/visproj/proofkit-gpt-oss-20b-lora) | gpt-oss-20b LoRA — the distillation **teacher** |
|
||
| [`visproj/proofkit-distilled-qwen0.5b`](https://huggingface.co/visproj/proofkit-distilled-qwen0.5b) | Qwen2.5-0.5B distilled from the teacher (merged) |
|
||
| [`visproj/proofkit-distilled-qwen0.5b-gguf`](https://huggingface.co/visproj/proofkit-distilled-qwen0.5b-gguf) | GGUF of the distilled student (llama.cpp — **served**) |
|
||
| [`visproj/proofkit-sft`](https://huggingface.co/datasets/visproj/proofkit-sft) | SFT dataset (synthetic, license-safe) |
|
||
| [`visproj/proofkit-distill-qwen0.5b`](https://huggingface.co/datasets/visproj/proofkit-distill-qwen0.5b) | Distillation dataset (teacher completions) |
|
||
|
||
### 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.
|