README.md

metadata

pretty_name: Amber Framework Knowledge Pack (demo)
license: mit
language:
  - en
task_categories:
  - question-answering
  - text-generation
size_categories:
  - n<1K
tags:
  - knowledge-pack
  - fine-tuning
  - lora
  - synthetic
  - crystal
  - amber
  - mlx
  - arxiv:2502.14502
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/valid.jsonl

Amber Framework Knowledge Pack (demo)

The demo knowledge pack dataset behind AgentC-Consulting/knowledge-packs: teach a small local model the Amber web framework (Crystal), and measure whether it learned anything with a before/after eval harness.

A knowledge pack compiles a body of expertise into curated sources, schema-validated generated training JSONL, a contamination-guarded held-out eval set, and a JSON manifest. This repo ships the exact training data and the two 50-item eval sets used for the measured run on record (live-run/RESULTS.md).

What's here

data/train.jsonl          471 training pairs (mlx chat format: {"messages": [...]})
data/valid.jsonl          24 validation pairs
eval/memorization.jsonl   50 items (30 MC + 20 recall) — trained-fact probe
eval/heldout_sources.jsonl 50 items (30 MC + 20 recall) — overclaiming control
eval/stats.json           generation stats (unit counts, split salt, seeds)

Generated deterministically (no model, no API) from amberframework/docs @ 3648105: 336 heading-chunked source units, split before generation by sha256(unit_id + "kp-amber-v1") into 262 train / 74 eval units. Cloze facts are sentences containing exactly one blankable inline-code term; each train fact yields 3 pairs (2 cloze phrasings + 1 multiple-choice phrasing).

The two eval sets — read this before quoting numbers

• memorization probes facts that ARE in the training data through question templates that never appear in training. It is a trained-fact probe measuring fact injection, not a source-held-out eval.
• heldout_sources draws facts only from the 74 eval-split units, which contributed zero training examples (unit-ID exclusion + dedup of blanked sentences into train). It is the leakage/overclaiming control: a memorization-only adapter should NOT improve here. Semantic restatements of a fact worded differently inside a train unit cannot be fully excluded by these guards.

Measured on mlx-community/SmolLM-135M-Instruct-4bit (M1 Max, LoRA, lr 1e-4, 2000 iterations): recall of trained facts 1/20 → 11/20 at temp 0; the held-out-source control stayed flat (9/50 → 7/50). Per-item reports (question, gold, full completion, verdict) are checked in at live-run/reports/. The eval questions are synthetic and not hand-reviewed; treat all numbers as lab notes, and run the harness yourself — the generator and scorer are in the repo (live-run/).

The trained adapter these numbers were measured on is published at crimson-knight/SmolLM-135M-Instruct-4bit-amber-lora.

Quality ceiling (stated plainly)

Deterministic template generation tests memorization and lookup-like recall only — no conceptual synthesis, no troubleshooting judgment. The 135M 4-bit base model is floor-level capability; the signal is in the recall probes, not multiple choice.

Related reading on fact injection via LoRA: arXiv:2502.14502 (How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?).

Sources & license

Condensed from the Amber Framework documentation (amberframework/docs) with attribution. The Amber framework itself is MIT; its docs repository carries no explicit LICENSE file (as of commit 3648105). The generated pairs, eval sets and tooling in this repo are MIT.

---

Built by the team at AgentC Consulting — https://agentc.consulting?ref=kp-dataset