Our model: conversation-v1, a generative research pilot.
362,614,656 unique parameters; 5,264 native token IDs.
Base: Qwen/Qwen2.5-0.5B
Revision: 060db6499f32faf8b98477b0a26969ef7d8b9987
Source: https://huggingface.co/Qwen/Qwen2.5-0.5B
Copyright and Apache-2.0 license: retained in MODEL-LICENSE.txt.

Modifications: emoji vocabulary transplant, eight additional framing/control rows,
embedding alignment and rank-16 LoRA adaptation on 1,888 synthetic/authored conversations.
Calibration selected adaptation epoch 3; adapters merged into the inference checkpoint.
The upstream authors do not endorse this independent experiment.

Inference runs on our server, with native emoji input and actual conversation history.
Greedy generation, reserved framing-token suppression only, 1,024 history tokens,
24 generated-token budget. No response lookup, appraisal routing or semantic correction.
The application does not persist conversation content. Infrastructure records request metadata.

31/32 held-out 18-turn curriculum conversations passed the authored scorer; task-specific
rules passed 32/32. Shared templates and synthetic labels limit what this establishes.
Six broader probes show irrelevant associations, missed negation and weak repair.
Zero independent human reviews. This is not a reliable unrestricted chatbot or proof
that the pretrained backbone retains its general knowledge in the emoji language.

The separate v27 logic sandbox is deterministic. EmojiLM remains an independent third-party
emoji-to-English interpretation tool; see emojilm/NOTICE.txt. No affiliation or endorsement.

Interactive controls: System 1 uses the existing single-pass candidate-scoring runtime with severity, binary, polarity, impact and tradeoff sets. The dialogue checkpoint has not been validated or calibrated for these classification tasks. System 2 uses autoregressive generation, optionally with sampling, temperature, top-p and a bounded output length; it is not a separately trained reasoning engine. Neutral defaults retain the original greedy reply behavior.

The expression preview is a separate 353-parameter ranker trained to approximate an authored preference function over eight styles. Warmth, energy, playfulness, flirtation, expressiveness and strength affect only this preview. Eligibility uses explicit context rules, not inferred chat meaning. The preview does not steer the conversational decoder.