Model card: mirror the astrea.altworld.io landing page (Altworldbench, samples, availability) + fine print

README.md

@@ -12,175 +12,259 @@ tags:
 - altworld
 ---
 
-# Astrea R8 Chat 9B
-
-Astrea R8 Chat is Altworld.io's 9B-parameter Qwen3.5 fine-tune for ordinary
-conversation, grounded instruction following, and creative prose. It is the
-native-chat release of the Astrea R8 line.
-
-**Try it:** [chat.altworld.io](https://chat.altworld.io) (free daily turns,
-no account needed) · [api.altworld.io](https://api.altworld.io)
-(OpenAI-compatible API) · [astrea.altworld.io](https://astrea.altworld.io)
-(docs and release notes)
-
-The bundled template is deliberately usable without a system prompt. When
-`enable_thinking` is absent or false, it inserts Qwen's closed empty thinking
-block and starts the visible answer immediately. Explicit thinking remains
-available with `enable_thinking=True`, but the release was evaluated primarily
-in non-thinking mode.
-
-## Intended use
-
-- General conversation and practical questions
-- Rewriting and instruction-following tasks
-- Long-form narrative and creative-writing assistance
-- Research on small open chat and creative-writing models
-
-This release is text-first. It inherits Qwen3.5's multimodal architecture and
-processor files, but Astrea's fine-tuning and reported evaluation did not
-establish independent vision performance.
-
-## Evaluation
-
-| Metric | Result | Notes |
-|---|---:|---|
-| Formal pooled narrative Elo | 1113.7 (1089.1–1140.9) | 794 comparisons; 95% bootstrap interval |
-| Held-out native-chat pass | 92.27% | 207 bare-chat prompts |
-| Ordinary-chat pass | 92.16% | No system prompt |
-| Accidental-fiction rate | 0.00% | Held-out chat suite |
-| Unseen identity contract | 20/20 | Paraphrases not in identity SFT |
-| Grounding | 73.44 | Frozen 50-prompt suite |
-| Slop composite | 2.522 | Frozen 120-prompt narrative suite; lower is better |
-| Length / POV / tense adherence | 92.50% / 93.33% / 99.17% | Frozen narrative suite |
-
-The 1113.7 Elo point estimate ranked first in the frozen seven-model pool. Its
-confidence interval overlaps Astrea R8 and Rocinante-X-12B-v1, so the supported
-claim is **#1 on point estimate and statistical parity with the top cluster**,
-not a statistically separated win.
-
-Generations were judged by a mixture of deterministic checks and LLM judges.
-The formal narrative pool used GLM-5p1 and DeepSeek-V4-Flash pairwise judgments,
-with Kimi-K2.6 adjudicating disagreements. See the repository release audit for
-the exact protocol and frozen artifacts.
-
-## Transformers usage
-
-Use a recent Transformers version with Qwen3.5 support.
-
-```python
-from transformers import AutoModelForMultimodalLM, AutoProcessor
-
-model_id = "Altworld/Astrea-R8-Chat-9B"
-processor = AutoProcessor.from_pretrained(model_id)
-model = AutoModelForMultimodalLM.from_pretrained(
-    model_id,
-    dtype="auto",
-    device_map="auto",
-)
-
-messages = [{"role": "user", "content": "Good afternoon."}]
-inputs = processor.apply_chat_template(
-    messages,
-    add_generation_prompt=True,
-    tokenize=True,
-    return_dict=True,
-    return_tensors="pt",
-    # Optional because this release defaults to false:
-    enable_thinking=False,
-).to(model.device)
-
-output = model.generate(
-    **inputs,
-    max_new_tokens=256,
-    do_sample=True,
-    temperature=0.2,
-    top_p=0.95,
-)
-new_tokens = output[0, inputs["input_ids"].shape[-1]:]
-print(processor.decode(new_tokens, skip_special_tokens=True))
+# Altworld Astrea: The New Frontier of Creative Writing
+
+**[Try Astrea →](https://chat.altworld.io)** · **[Get an API key](https://altworld.io/developer)** · **[Read the docs →](https://astrea.altworld.io/docs/)**
+
+Today, we are introducing **Altworld Astrea** — our first openly released
+model. Astrea is a compact 9B-parameter model built for one thing: writing
+that feels written. It sounds human, it keeps the story straight, and it
+beats bigger models at both. You say something, it answers back.
+
+**Altworldbench — overall writing quality.** Elo from 794 blind matchups on
+our release pool — both rivals are bigger models built for creative writing:
+
+| | Elo |
+|---|---:|
+| **Astrea 9B** | **1114** |
+| Rocinante-X 12B | 1097 |
+| Wayfarer-2 12B | 1045 |
+
+Astrea is available today on [chat.altworld.io](https://chat.altworld.io),
+and the full model weights are right here under Apache-2.0, so anyone can
+download it and run it locally. An OpenAI-compatible
+[API](https://astrea.altworld.io/docs/) is open with simple prepaid pricing.
+
+## What makes Astrea different
+
+### It sounds like a person
+
+Most AI prose gives itself away in the first paragraph: the stock phrases,
+the "it wasn't X — it was Y", the em-dash pileups, every reply opening the
+same way. We count those tells. Astrea scores **2.5** — roughly two and a
+half tells slipping into every thousand words. Every other model we tested
+leaves between 8 and 19 per thousand.
+
+### It understands the scene, not just the sentence
+
+Judged blind, story against story, Astrea wins on the things that make
+writing actually work: characters whose motivations hold together, a detail
+planted early paying off later, the tone you asked for instead of generic
+fantasy voice. That depth is what puts it ahead of models three billion
+parameters bigger.
+
+### It stays grounded
+
+If the lantern went out, it stays out. If your sword was taken, it doesn't
+reappear in your hand two scenes later. That's grounding — remembering what's
+true in the story — and creative models are notoriously bad at it. Astrea
+keeps **73 of 100** scenes straight. The strongest 12B prose model manages
+**44**.
+
+### All of it at 9B
+
+Every rival on our board is both bigger than Astrea *and* purpose-built for
+creative writing. Astrea isn't beating generalists — it's beating the
+specialists, with three billion fewer parameters. That also means it runs
+faster, serves cheaper, and fits on hardware the bigger models can't.
+
+## Altworldbench
+
+We test every release on **Altworldbench**, our own creative-writing
+benchmark. Same prompts and settings for every model, every story judged
+blind.
+
+**Head-to-head wins** — share of blind matchups Astrea wins:
+
+| | wins |
+|---|---:|
+| vs Rocinante-X | 52% |
+| vs Wayfarer-2 | 65% |
+
+**AI-sounding writing** — lower is better:
+
+| | score |
+|---|---:|
+| **Astrea 9B** | **2.5** |
+| Rocinante-X | 8.1 |
+| Wayfarer-2 | 8.3 |
+| Muse | 8.8 |
+| Violet Lotus | 12.8 |
+
+**Keeps the story straight** — out of 100:
+
+| | score |
+|---|---:|
+| **Astrea 9B** | **73** |
+| Violet Lotus | 68 |
+| Muse | 64 |
+| Wayfarer-2 | 61 |
+| Rocinante-X | 44 |
+
+**Writes the length you ask for** — % of the time:
+
+| | % |
+|---|---:|
+| **Astrea 9B** | **93** |
+| Rocinante-X | 88 |
+| Violet Lotus | 83 |
+| Wayfarer-2 | 78 |
+| Muse | 25 |
+
+All rivals are open 12B creative-writing models.
+
+## See for yourself
+
+Two samples from our test set — unedited, exactly as Astrea wrote them.
+
+> **SCENE:** The rescue party returns with empty stretchers and one boot.
+> **ACTION:** I ask the search captain to tell the families myself.
+
+> The captain sets the boot on the table. It is a heavy boot, the kind the
+> welders wear, with the steel cap dented in and the laces frayed. The
+> stretchers are empty now, returned to the hangar. The rescue party came
+> back in four, and one of them did not return.
+>
+> The families are waiting in the corridor. They have not eaten in days.
+> The ration scrip is worthless now that half the station is gone, and the
+> spin-section has slowed to a crawl. The air smells of recycled carbon and
+> old sweat. You do not speak of the lost. You do not speak of the dead.
+>
+> The captain turns to you. He does not speak. The boot is on the table. He
+> steps back.
+>
+> You stand. You take the boot. You walk into the corridor. The families
+> look at you. There is nothing to say. You set the boot on the floor and
+> say, "Your children did not come back."
+
+> **SCENE:** Your partner swore your cargo of walrus ivory to two different
+> buyers.
+> **ACTION:** I go to the lesser buyer first and offer him the truth and a
+> discount.
+
+> The lesser buyer squints at you through a greasy half-door in Uppsala's
+> timber shed. The air smells of pine pitch and old fish.
+>
+> He does not lie. The other party is already buying that ivory. Three
+> measures from Helsingjar, not yours. You tell him the truth, offer him a
+> discount of two dirhams. He laughs, spits. "The other man has my seal. He
+> paid me in hacksilver and a vow over the sacred waters. You offer scraps,
+> and I must burn them for luck."
+>
+> He turns away. The floorboards are wet. Frost lies thin on Lake Mälaren
+> beyond the sheds, and the ice does not hold for a barge. The cargo
+> remains in the gloom: twenty tusks, the colour of a seal's eye, unsworn
+> to you now.
+
+## Availability
+
+- **Talk to Astrea** — chat with it now at
+  [chat.altworld.io](https://chat.altworld.io) — no account needed for your
+  first turns.
+- **Build with the API** — an OpenAI-compatible endpoint at
+  `api.altworld.io` — create a key and load credits at
+  [altworld.io/developer](https://altworld.io/developer), then
+  [start in the docs](https://astrea.altworld.io/docs/).
+- **Play Astrea** — Astrea powers the voices of
+  [Altworld](https://altworld.io), our life simulator and story sandbox.
+- **Download the weights** — this repository: Apache-2.0, BF16 safetensors.
+
+## Build with the API
+
+The Astrea API is **OpenAI-compatible**: if your code already talks to an
+OpenAI-style endpoint, point it at our base URL, swap the model name, and it
+works — streaming included.
+
+```bash
+curl https://api.altworld.io/v1/chat/completions \
+  -H "Authorization: Bearer $ASTREA_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "astrea",
+    "messages": [{"role": "user", "content": "Open a story in a lighthouse, first person, no cliches."}],
+    "stream": true
+  }'
+```
+
+## Model details
+
+Altworld Astrea (release name Astrea-R8-Chat-9B) was trained on top of the
+Qwen3.5-9B base model using multi-round LoRA supervised fine-tuning,
+alternating conversational and narrative rounds and checking every round
+against Altworldbench. The released checkpoint is the merged BF16 model.
+
+The chat template works bare: no system prompt, no flags, it just answers —
+with no visible reasoning block. Explicit thinking is available per-request
+with `enable_thinking=True`.
+
+Recommended sampling:
+
+```json
+"temperature": 0.8,
+"min_p": 0.025,
+"repetition_penalty": 1.08
 ```
 
-The exact recommended runtime class may change as Qwen3.5 support lands in
-stable Transformers/vLLM releases. Follow the base model's official runtime
-instructions and pin a tested revision in production.
+Use `temperature 0.2` for factual chat.
+
+## Limitations
 
-## Serving with vLLM (production-validated)
+- Like any compact model, Astrea can miss nuance or state things confidently
+  and wrongly. Verify anything high-stakes independently.
+- English dominates its training and evaluation. Other languages will work,
+  but with weaker guarantees.
+- This release is text-first: the architecture inherits vision support, but
+  visual behavior was not part of the release evaluation.
+- Astrea is not certified for medical, legal, financial, or other
+  high-stakes use.
 
-This checkpoint powers Altworld's own endpoints on vLLM 0.25, validated on
-an A40 and on 2× RTX 3090 (`--tensor-parallel-size 2`): identity contract,
-no accidental fiction, streaming, stop tokens, and seed-fixed sampling
-reproducibility all pass on this engine.
+We plan to keep improving and openly releasing Astrea models — tell us what
+to fix. The fastest way to reach us is the Altworld Discord, or open a
+Discussion on this repository.
+
+---
+
+## The fine print
+
+**Evaluation protocol.** The pooled Elo is 1113.7 with a 95% bootstrap
+interval of 1089.1–1140.9 on a frozen seven-model pool (794 comparisons for
+Astrea's entry). The interval overlaps Astrea R8 and Rocinante-X-12B-v1, so
+the supported claim is #1 on point estimate and statistical parity with the
+top cluster — not a statistically separated win. Pairwise judgments were
+made blind by GLM-5p1 and DeepSeek-V4-Flash, with Kimi-K2.6 adjudicating
+disagreements; grounding, slop, length/POV/tense, and bare-chat checks
+combine deterministic rules with LLM judging. Separately from the narrative
+pool: 92.27% pass on 207 held-out bare-chat prompts, 0.00% accidental
+fiction, 20/20 on unseen identity paraphrases. One constructed multi-turn
+memory probe answered indirectly at temperature 0.2 while fine at 0.8 —
+test your own dialogue distributions.
+
+**Running it locally.** Validated in production on vLLM 0.25 (A40, and
+2×RTX 3090 with `--tensor-parallel-size 2` — set `NCCL_P2P_DISABLE=1` on
+consumer dual-GPU boards):
 
 ```bash
 vllm serve Altworld/Astrea-R8-Chat-9B \
-  --dtype bfloat16 \
-  --max-model-len 32768 \
-  --gpu-memory-utilization 0.90
+  --dtype bfloat16 --max-model-len 32768 --gpu-memory-utilization 0.90
 ```
 
-- The base architecture supports 262,144 tokens of context natively; serve
-  whatever your VRAM affords (32k fits comfortably on 2×24 GB alongside the
-  BF16 weights — the hybrid linear-attention layers keep the KV footprint
-  small).
-- On consumer dual-GPU boards without P2P, set `NCCL_P2P_DISABLE=1` or
-  tensor-parallel start-up can deadlock.
-- Sampling profiles we serve in production: factual/plain
-  `temperature 0.2, top_p 0.95`; conversational/creative
-  `temperature 0.8, min_p 0.025, repetition_penalty 1.08`.
-- GGUF: llama.cpp supports the Qwen3.5 family; official quants of this
-  fine-tune are planned — watch this org.
-
-## Serving notes
-
-- Do not add a fictional roleplay system message for ordinary chat.
-- Keep thinking disabled for the default user experience.
-- Start factual/basic chat around temperature 0.2; creative generation was
-  evaluated around temperature 0.8.
-- Preserve both Qwen stop tokens from `generation_config.json`.
-- A host may quantize the checkpoint to FP8. Re-run the smoke and narrative
-  checks for that host; FP8 kernels are not interchangeable quality guarantees.
-
-## Known limitations
-
-- Like any 9B model, Astrea can miss nuance, hallucinate, or follow ambiguous
-  instructions poorly. Verify high-stakes claims independently.
-- One constructed multi-turn memory probe (“Mira” plus adopting a cat) exposed
-  low-temperature indirectness/confusion even though the facts remained in
-  context; higher temperature answered it directly. Applications should test
-  their own dialogue distributions.
-- Engine portability: the production engine (vLLM 0.25, BF16) passes the
-  full release gate battery, and the Fireworks source runtime matched it. A
-  local MLX-LM smoke run was weaker on identity and user/model-name
-  separation at temperature 0.2 despite numerically matched logits — treat
-  runtime, kernels, and sampling implementation as part of the artifact and
-  validate your exact engine before depending on it.
-- Identity behavior is a trained convention, not evidence of consciousness or
-  an independent person.
-- English dominates the fine-tuning and evaluation evidence.
-- Safety behavior was not certified for medical, legal, financial, or other
-  high-stakes deployment.
-- Vision behavior was not part of the release evaluation.
-
-## Lineage and license
-
-Base model: [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B),
-licensed under Apache License 2.0. This derivative is released under the same
-license. See `NOTICE` for attribution and modification information.
-
-Fine-tuning preference data used gold responses generated by MIT-licensed
-open-weight models (GLM-5.2 and DeepSeek-V4-Flash) under deterministic task
-checks and an independent judge gate, with on-policy generations from the
-Astrea lineage as rejected sides. No closed-source model outputs were used
-at any stage.
-
-## Acknowledgements
-
-Early-stage data ancestry draws on openly licensed community datasets:
-Athanorlite-DPO (CC-BY-4.0), jondurbin/gutenberg-dpo-v0.1 (CC-BY-4.0),
+The base architecture natively supports 262k context; the hybrid
+linear-attention layers keep the KV cache small, so serve what your VRAM
+affords. With Transformers, load via `AutoProcessor` +
+`AutoModelForMultimodalLM` and apply the bundled chat template
+(`enable_thinking` defaults off). An MLX-LM smoke run behaved worse than
+vLLM at temperature 0.2 despite numerically matched logits — validate your
+exact engine. GGUF quants are planned; llama.cpp supports the Qwen3.5
+family.
+
+**Lineage and license.** Base model
+[Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B), Apache-2.0; this
+derivative is released under the same license — see `NOTICE`. Fine-tuning
+preference data used gold responses from MIT-licensed open-weight models
+(GLM-5.2, DeepSeek-V4-Flash) with on-policy rejected sides; no closed-model
+outputs anywhere. Early-stage data ancestry credits: Athanorlite-DPO
+(CC-BY-4.0), jondurbin/gutenberg-dpo-v0.1 (CC-BY-4.0),
 nbeerbower/gutenberg2-dpo and gutenberg-moderne-dpo (Apache-2.0),
-Pageshift-Entertainment/LongPage (public-domain text, CC-BY-4.0
-annotations), allenai/tulu-3-sft-mixture (ODC-BY, license-filtered slices),
-and nothingiisreal/Reddit-Dirty-And-WritingPrompts (Apache-2.0, SFW file
-only). Thanks to the Qwen team for the Apache-2.0 base model.
-
-Questions or regressions: open a Discussion on this repository.
+Pageshift-Entertainment/LongPage, allenai/tulu-3-sft-mixture (ODC-BY,
+license-filtered), nothingiisreal/Reddit-Dirty-And-WritingPrompts
+(Apache-2.0, SFW file only). Thanks to the Qwen team for the base model.