Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,98 @@
+---
+license: other
+language: [en]
+task_categories: [text-classification, text-generation]
+tags:
+  - verdict-engine
+  - build-small-hackathon
+  - distillation
+  - triage
+  - llm-as-a-judge
+  - synthetic
+size_categories: [n<1K]
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: train_v2.jsonl
+      - split: validation
+        path: eval_v2.jsonl
+---
+
+# Verdict Engine — SFT v2 (Sonnet-4.6 distilled)
+
+The supervised fine-tuning dataset behind every model in the
+[Verdict Engine bake-off](https://github.com/setuc/verdict-engine): four LoRA
+fine-tunes (`hqt2yotoz/verdict-engine-{qwen3-4b-2507,qwen3-8b,smollm3-3b,qwen2.5-7b}-lora`)
+that all beat their own base on a blind panel. This is the **v2 / "true-distillation"**
+dataset — the one whose labels come from a genuinely stronger teacher (Claude Sonnet 4.6),
+not from the small model labeling itself.
+
+## What it teaches
+
+The Verdict Engine has two faces, and this dataset trains both with one chat schema family
+(`messages`: `system` / `user` / `assistant`, where the assistant turn is a single JSON object):
+
+| Task | Examples (train / val) | Persona(s) | Assistant JSON |
+|---|---|---|---|
+| **Triage** | 92 / 12 | The Librarian | `label` ∈ {read-now, skim, skip, archive}, `novelty_score` ∈ [0,1], `one_line`, `tags` (2–5), `reason` |
+| **Critic panel** | 376 / 40 | Reviewer #2 · The Over-Eager Intern · The Weary Professor · The Hype Beast | `verdict`, `roast`, `score` |
+| **Total** | **468 / 52** | | |
+
+### Triage label distribution (train split)
+`skip` 58 · `read-now` 16 · `skim` 14 · `archive` 4.
+
+This is **deliberately skip-heavy, and that is the point.** Most of the daily model/dataset/paper
+firehose is genuinely skippable; a calibrated triager should say so. The single most important
+finding behind these models is that base models *over-praise* (almost everything → `read-now` with
+inflated novelty), while a model distilled on this distribution learns to **discriminate** — which
+is what wins the blind eval. A balanced label set would have taught the wrong thing.
+
+## Provenance — true distillation, not self-distillation
+
+Items are real entries from the Hugging Face firehose (newest models & datasets via
+`huggingface_hub.list_models` / `list_datasets`, plus the HF Daily Papers API) and a small set of
+hand-seeded items. Each item's gold label was produced by **Claude Sonnet 4.6** through an AI
+gateway acting as the teacher (`finetune/gateway_teacher.py`), then formatted into chat SFT records
+(`finetune/gen_dataset.py`).
+
+This is the methodological core of the project. An earlier **v1** dataset was labeled mostly by the
+*same* 7B model being fine-tuned — self-distillation — and the resulting fine-tune **lost** its
+blind eval 6/3/3 (a model can't teach itself a skill it lacks). Re-labeling with a true stronger
+teacher (this v2 set) flipped that to a **12–0 sweep** on the original 7B, and to clean wins for
+all four bases in the multi-model bake-off. See
+[`docs/TRAINING.md`](https://github.com/setuc/verdict-engine/blob/main/docs/TRAINING.md).
+
+## Files
+- `train_v2.jsonl` — 468 records, one JSON object per line, key `messages`.
+- `eval_v2.jsonl` — 52 held-out records, same schema.
+
+The held-out **bake-off eval items** (the fixed 10 the blind panel scored every model on) and the
+blind/keyed eval pairs live in the GitHub repo under `finetune/bakeoff/`.
+
+## How it was used
+Identical recipe across all bake-off models — only the base varies:
+LoRA `r=16, α=16`, targets `q,k,v,o,gate,up,down`; **bf16 LoRA via PEFT + TRL** (not 4-bit — the
+Unsloth 4-bit path corrupted adapter weights for these bases; see TRAINING.md §4); trained on each
+model's **native chat template**; 3 epochs on an AWS A10G. Reproduce:
+
+```bash
+python finetune/train_lora_mps.py \
+  --model Qwen/Qwen3-4B-Instruct-2507 \
+  --train train_v2.jsonl --eval eval_v2.jsonl \
+  --out out_qwen3 --epochs 3
+```
+
+## Honest limitations
+- **Synthetic labels.** Gold verdicts are Sonnet-4.6 judgments, so the dataset inherits that
+  teacher's taste and blind spots. It is a *narrow, in-domain classifier* signal — good for this one
+  triage/critique schema, not a general-purpose judge (cf. Huang et al., ACL 2025 Findings,
+  *"A Fine-tuned Judge Model is not a General Substitute for GPT-4"*).
+- **Small + imbalanced.** 520 examples, skip-heavy triage labels, critic-mode dominant. Treat as a
+  task-shaping SFT set, not a benchmark.
+- **Item text is public metadata** (HF card snippets, arXiv-style abstracts, titles) captured in
+  June 2026; original works retain their own licenses. The *labels/personas* are model-generated.
+  `license: other` reflects this mixed provenance — released for research/reproducibility of the
+  Verdict Engine experiment.
+
+— Built for the Hugging Face **Build Small** hackathon. Code: github.com/setuc/verdict-engine