---
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
library_name: transformers
tags: [le-harnais, distillation, agentic-coding, scaffold, hero]
---

# le-harnais / ft-ornith-scaffold-3b

A **scaffold-generation** student: given a coding/agent task it emits a short numbered plan of
subgoals, each tagged with how to **verify** it. Distilled (plain full-FT SFT) from the
`ornith:35b` teacher's self-scaffolds, so a small local model can produce the structured plans
that le-harnais's `lh harness self-scaffold` step executes and verifies.

- **Base model:** [`meta-llama/Llama-3.2-3B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) — **Built with Llama**; Llama 3.2 Community License applies.
- **Class:** `hero`
- **Teacher:** `ornith:9b`/`ornith:35b` family (MIT) — scaffolds generated by `ornith:35b`.
- **Training data:** 153 `(task → verify-tagged numbered plan)` pairs from the teacher (38 held-out).
- **Headline (held-out, n=38, vs the base):** token-F1 vs teacher **0.295** (base 0.170, **+0.125**);
  scaffold-structure score **0.731** (base 0.577, **+0.153**).

## Why it exists

le-harnais's thesis is *don't trust raw model text — generate, verify, repair*. Ornith's
"self-improving" RL learns to propose its own search scaffold; we can't reproduce that RL on one
GPU, but we **can distill the scaffolding behavior** into a cheap local model. This student is
the result: it drives the inference-time `--self-scaffold` path (which on a weak policy halved
escalation in an A/B), with no RL and no remote model.

## Reproduce

```sh
# 1. corpus from the teacher, 2. full-FT distill, 3. held-out bench
python tools/gen_ornith_scaffold_corpus.py --n 160 --eval 40 --out datasets/ornith_scaffold
BASE=meta-llama/Llama-3.2-3B-Instruct OUT=refs/llm-jepa/ft-ornith-scaffold-3b \
  tools/distill_ornith_scaffold.sh
python tools/bench_scaffold.py --student refs/llm-jepa/ft-ornith-scaffold-3b \
  --base meta-llama/Llama-3.2-3B-Instruct --eval datasets/ornith_scaffold_eval.jsonl
```

Full context: `docs/ornith-self-improving.md` (D1–D4) and `docs/local-coding-agents-bench.md`
in the [le-harnais] repo. Serve with `transformers` or le-harnais `lh-serve`/candle.

> Distilling a *teacher's scaffolding skill* into a small student: structure and plan quality
> transfer; this is behavior distillation, not the teacher's RL.