A-Level Physics CIE (9702) Answer Template Generator#

Generate structured answer templates for Cambridge International A-Level Physics (9702) questions using a fine-tuned Qwen3-4B model with LoRA adapters trained on 1652 real past papers.

Skill contract (runtime vs optional tooling)#

This section clarifies what runs for normal skill use vs what exists only for dataset rebuild / retraining, so automated reviewers (e.g. OpenClaw) and humans can align expectations with the code.

Primary path — inferenceOptional — training / data pipelineEntrypointsskill/scripts/inference.py, generate_template / generate_template_verified in that modulescraper/*, scripts/build_sft.py, scripts/run_full_pipeline.py, scripts/train.sh, etc.Remote APIsNone for generationDeepSeek API when --teacher deepseek or full pipeline teacher mode ( DEEPSEEK_API_KEY)Web / HTTPHugging Face (typical) to download base model Qwen/Qwen3-4B-MLX-4bit on first run; no user question leaves your machine as HTTP payloadcie.fraft.org when running the scraper; HF again for training stack as configuredSecrets****No DEEPSEEK_API_KEY required by inferenceDEEPSEEK_API_KEY only if you regenerate SFT via DeepSeek

Inference does not scrape past papers, does not call DeepSeek, and does not exfiltrate prompts to a third-party LLM API. Maintainer scripts may; they are separate.

Full detail: SECURITY.md in the repository root.

Mandatory rule for the orchestrator (plain-text math)#

When you produce any final answer, template, or paraphrase for the user—whether you ran skill/scripts/inference.py or answered from general knowledge—you must:

1. Write formulae in plain text (e.g. v² = u² + 2as, E = hf, λ = h/p, P = IV).
2. Never wrap math in $...$, $$...$$, \(...\), \[...\], or similar TeX delimiters. Raw $$ is unreadable for users in Clawhub/OpenClaw-style clients.
3. If tool output still contains stray $ signs, strip or rewrite those segments into plain text before showing them to the user.

Local inference already applies the same rule via its system prompt and post-processing; the orchestrator must follow it even when not calling the script.

Quick Start#

Run inference on a physics question:

bash

python skill/scripts/inference.py "Define specific heat capacity."

Or in Python:

python

from skill.scripts.inference import generate_template
result = generate_template("Calculate the maximum height reached by a ball thrown upward at 20 m/s.")
print(result)

Output Format#

The model produces structured answer templates:

• Question type — calculation / definition / explain / describe / derive / analyse / practical
• Given — quantities and conditions from the question
• Required — what the student must find or state
• Formulae / principles — relevant equations and physics laws
• Answer frame — numbered step-by-step approach
• Check — unit/sign/direction/significant-figure verification

Display note (Clawhub / chat clients — applies to orchestrator and model): Present equations in plain text (ASCII and Unicode, e.g. v², λ, ×, fractions with /). Do not use LaTeX delimiters ( $, $$, \(…\), \[…\]) in final user-facing output — many clients do not render math, so those tokens look garbled. The inference script enforces this with a system prompt and post-processing when you run it; if you answer without the script, you must still follow this rule.

Model Details#

• Base model: Qwen/Qwen3-4B-MLX-4bit
• Adapter: LoRA rank 8, 16 layers, trained 1000 iterations
• Training data: 414 question–template pairs from 9702 Papers 2/4/5 (2001–2025), templates generated by DeepSeek with mark-scheme context
• Peak memory: 4 GB (runs on any 8GB+ Apple Silicon Mac)

Retraining#

To retrain or extend with more data:

bash

python scripts/run_full_pipeline.py --teacher deepseek

See skill/references/training.md for the full pipeline details.

Adversarial Robustness Evaluation#

Test the model's robustness using three physics-adapted attack strategies from Xie et al. (2024):

bash

python skill/scripts/adversarial_eval.py
python skill/scripts/adversarial_eval.py --strategies numeric --variants 5 --max-questions 10

Reports OA (Original Accuracy), AA (Adversarial Accuracy), and ASR (Attack Success Rate) per strategy.

References#

• skill/references/training.md — Full scraping, extraction, SFT, and training pipeline
• skill/references/answer_template_format.md — Detailed output format specification
• skill/scripts/inference.py — Standalone inference script
• skill/scripts/adversarial_eval.py — Adversarial robustness evaluation (numeric perturbation, context swap, question-type adversarial)
• SECURITY.md — Network, secrets, and trust boundaries