ELIZA Qwen3 0.6B

This is a narrow behavioural fine-tune of Qwen/Qwen3-0.6B-Base. It produces short, uppercase, reflective replies in the style associated with the classic ELIZA DOCTOR script. It was trained locally with LoRA on synthetic conversations from the repository's deterministic rule oracle.

Intended use

• Historical and educational demonstrations of early chatbot behaviour.
• Studying whether a small base model can imitate a deterministic dialogue policy.
• Local, low-resource text generation experiments.

This model is not a therapist, mental-health service, factual assistant, or general-purpose chat model. Its characteristic behaviour is to reflect and deflect rather than answer questions.

Training data

The dataset contains 30,000 training conversations plus 2,000 validation and 2,000 test conversations. It is generated locally and contains no private user conversations or third-party chat corpus. Test and validation sets use complete wording frames and slot vocabularies absent from training. The target behaviours include viewpoint reflection, feelings, identity statements, inability, wants, family, dreams, computers, reasons, questions, universals, uncertainty, affirmation, negation, greetings, apologies, fallback, goodbye, and multi-turn memory recall.

Only final assistant tokens receive loss; prior turns are context. The adapter targets all linear attention and MLP projections with rank 16 LoRA.

A 4,000-example corrective pass followed the initial run: 2,000 newly generated paraphrases covering held-out failure categories and 2,000 original examples for replay. The selected corrective checkpoint improved validation loss from 1.604 to 1.571.

Limitations

• The model imitates DOCTOR-style behaviour; it is not a bit-for-bit restoration of a particular historical implementation or 1980s port.
• The training grammar is synthetic and deliberately narrow.
• Reflective responses can appear empathetic without implying comprehension.
• It may respond inappropriately to crises because historical ELIZA behaviour is not a modern safety policy. Deployments should handle safety outside the model.
• Uppercase output is intentional terminal-era styling.

Evaluation

Aggregate results are written to artifacts/eliza-qwen3-0.6b/eval/report.json and per-example generations to predictions.jsonl. Because several oracle responses are valid for a prompt, evaluation combines response-shape properties, keyword rule signals, word overlap, and manual transcript inspection rather than relying on exact match alone.

On 500 held-out examples from complete wording frames and slot vocabularies not present in training:

Metric	Result
Non-empty response	100%
Concise response (1–24 words)	100%
Uppercase terminal style	100%
No modern-assistant phrase	100%
Coarse rule signal	99.6%
Oracle word-F1	0.461
Explicit multi-turn memory signal	100%

Under the same runtime protocol, the untouched base model scored 1% on coarse rule signal and 0.0006 word-F1 over a 100-example baseline sample.