this is efish other, an other. — live at efishother.com
一个只做对话的 PWA：它以一个人的"另一个"来回答，最多两句话。 它背后不是一层 API 外壳，而是一段研究：一个从零训练的 96M 中文 transformer、一套盲测与放行纪律、 几个被如实记录的否定结论、一份从本人写作里整理出来的人格契约，以及为它专门做的键盘、输入法和首页。
efish other is a chat-only PWA that answers as one person's other — a person, a memory, an efish (鳄鱼: the owner's alligator, which in this project is simply called efish) and a dialog box at once — in two sentences at most. This repository is the research behind it as much as the product: most of what is here is evidence about what did not work, kept because it decided what the product became.
1. A small model, trained from nothing. efishv1 is a 96M-parameter Chinese decoder trained
on one laptop, with its own tokenizer, packaged as a 4-bit browser artifact (5 shards,
48,267,968 bytes) with an exact runtime tokenizer and a static, no-backend runtime. It is released
under MIT and documented in MODEL_CARD.md and DATA_CARD.md. The runtime that shipped it is kept
byte-for-byte in archive/legacy_site/.
2. An evaluation discipline that was allowed to say no. Blind prompts split by session,
semantic, source and event family; assistant-response-only masking; no prompt-specific routing;
no editing of frozen thresholds to make a candidate pass; release gates that fail closed. Lower
loss was never accepted as proof of dialogue ability. The gates live in scripts/ and the
contracts in docs/.
3. Negative results, recorded with their numbers.
| Question | Test | Result |
|---|---|---|
| Can the 96M weights hold a dialogue after SFT? | R29B2M-R3: response-only SFT to 239,048 optimizer tokens | validation loss 4.82; 0 of 5 on all twelve generated-dialogue behaviour families → BLOCKED_DIALOGUE_QUALITY_WITH_EVIDENCE |
| Was the browser forward ever contextual? | code audit of the q4 runtime | no: attention ran over one token, without a KV cache — the shipped model never ran a correct multi-token forward in a browser |
| Can the small model steer a large one? | R29B2M-R4H-R3: temperature- and structure-controlled replay, 24 blind pairs, oracle signal | preference 50.0 % (needed ≥ 55), factual non-regression 79.2 % (needed ≥ 95), 7 unsupported facts (needed 0) → architecture rejected |
| Can it at least select among candidates? | R29P0 pairwise oracle, 20 live pairs | 0 % safe equivalent headroom → stopped before batch 2 |
See docs/R29B2M_R4H_R3_CONTROLLED_CAUSAL_CRITIC_HYBRID.md, docs/R29P0_EQUIVALENCE_PAIRWISE_ORACLE.md,
docs/r29/ and docs/r31/R31A0_DEEPSEEK_PRODUCT_PATH.md.
4. A persona that is compiled, not prompted into being. R30J set out a method for excavating a
person's judgment from their own writing (docs/R30J0_*, docs/R30J1*). The persona contract in
app/src/engine/persona_prompt.js is compiled from a private value-logic profile of the owner's
essays, poems, published posts and answered questions, and then tuned against live answers with
him. What that tuning established is part of the research:
- vivid source material placed in a prompt is lifted by the model as imagery — only judgments go in, never images, and a test keeps them out;
- the model invents memories and creeds unless told exactly what may be said — the owner's positions are used verbatim and nothing beyond them is asserted for him;
- his syntax is measurable (short comma-split phrases, about ten characters a clause) and is tuned with a tool that reports it (
npm --prefix app run ask); - an answer must not arrive in the same instant as the question — the rhythm of thinking and writing is part of the voice.
5. Where the small model stands today. It is not in the answer loop. Generation is DeepSeek-V4.1-Flash behind one reviewed relay; the 96M model and its training stack remain the project's own line of work. The open questions are a bilingual vector knowledge base over the owner's corpus with retrieval into the conversation, and retraining the small model for a design in which a large model exists — the previous one was trained before that was true.
| Home | one pinned stage; 25 pen strokes reused across four drawings (a tree, a lock, a cloud over two doors, a loop) that end as the chat window's border; words are spelled letter by letter and tied with half-circles; one gesture plays one act (app/src/ui/scene.js, pure geometry, tested) |
| Chat | a two-sided conversation, never more than two sentences (enforced in code while streaming), paced: the question lands, efish thinks, then writes |
| Keyboard | the system keyboard is never raised; a hand-drawn keyboard with a sentence-level pinyin IME (47k-key dictionary built from jieba + pypinyin, Viterbi composition, repair of neighbour-key slips, fuzzy initials and finals) and a board of preset questions |
| Memory | conversations stay on the device (IndexedDB) for thirty days; "new" starts another, 本地记忆 reopens old ones; about twenty exchanges travel as context |
| Key | held on the server; one Edge route, api/chat.js → app/server/chat_handler.js, pins the model, the token budget and the persona, checks the origin, rate-limits, never logs; its contract is written down in scripts/server_proxy_policy.mjs and enforced by the release gates |
| PWA | React 19 + Vite + Workbox; 1.9 MB install; OPPO Sans shipped unmodified under its own licence and loaded after first paint |
Round-by-round records: docs/r31/.
app/ the product (React PWA), its relay handler, tests and tools
api/chat.js the only server route: a one-line Edge entry into the relay
archive/legacy_site/ the static runtime that shipped efishv1 (q4 shards, worker, dashboard)
src/, training/ model lab, training and corpus tooling for the 96M line
docs/ contracts, methods and results for every round (R25 … R31)
scripts/, tests/ release gates, policies and their tests
MODEL_CARD.md DATA_CARD.md MODEL_LICENSE.md PRIVACY.md SECURITY.md
npm --prefix app ci
npm --prefix app run key # store a DeepSeek key locally (hidden input → app/.env.local)
npm --prefix app run dev # http://localhost:5190
npm --prefix app run ask -- "你是谁？" # question the relay from a terminal; --prod for the live siteChecks that describe the current product:
npm run test:app # engine, relay, scene, memory
npm run test:r31a0 # policies, incl. the relay contract
npm run check:static-local-product
npm run check:hybrid-lab-isolation
npm run check:app-budget # install ≤ 10 MB, service worker and manifest present
npm run test:r29b2m-r4h # the lab added no route; the only route is the reviewed relayThe R27–R30 suites that build or byte-freeze the legacy site describe the archived runtime and
are kept for the record (docs/r31/R31B2_CUTOVER.md).
Dai Pan / 潘岱 — visual artist, designer, photographer, printmaker and writer, who moved from fine art into computer science and NLP research. efish other is his other: the voice is compiled from his own writing and corrected by him line by line. See also the poems of Three Worlds at daipan.ink and IN_PRAISE_OF_TIME.
Raw private materials, the owner's value-logic profile and source writing, raw/clean/processed
corpus dumps, checkpoints, optimizer states, adapters, tokenizer-training artifacts, private
calibration data, hidden prompts and any API key. See PRIVACY.md.
The repository source code is MIT licensed. See LICENSE.
The committed R28M1 q4 browser model package, published here as efishv1, is also released under
MIT as a static model artifact package. See MODEL_LICENSE.md and MODEL_CARD.md.
This license grant covers committed source and committed public runtime model assets only. It does
not grant rights to uncommitted private files, raw source materials, ignored artifacts, raw
checkpoints, LoRA/adapters, tokenizer training artifacts, private calibration data, or external
third-party materials. OPPO Sans 4.0 is distributed unmodified under its own licence
(app/public/fonts/oppo-sans/LICENSE.txt).