flammenai/Phoenix-SFT-v1

Phoenix v1 — Memory Extraction Dataset

Synthetic SFT training data for Phoenix, a small (Qwen2.5-class) model that reads a window of chat messages between a user and an AI character ("Flame") and emits a structured JSON list of memorable facts the Flame should remember about the user going forward.

This is the v1 dataset shipped by flammen.ai for the memory layer that drives long-term continuity in async character conversations. Released as part of flammen.ai's open-source data strategy.

Task

Input: a conversation transcript like

User: Nina
Flame: Marisol

Conversation:
Nina: hey. been a rough week
Marisol: what happened?
Nina: my grandma passed last month and ive been like... kind of fine? and then today i found one of her recipe cards in a drawer and lost it
Marisol: oh nina. i'm so sorry.
Nina: yeah. she taught me to bake. her oatmeal cookies are like the only thing i can make from memory

Output: a JSON object with structured memories:

{
  "memories": [
    {"content": "Grandmother recently passed away", "category": "event"},
    {"content": "Was taught to bake by her grandmother", "category": "relation"},
    {"content": "Can make her grandmother's oatmeal cookies from memory", "category": "factual"},
    {"content": "Working through grief over her grandmother's death", "category": "emotion"}
  ]
}

Schema

Categories (closed enum):

• factual — concrete data (job, location, family structure, possessions)
• preference — likes/dislikes/values/habits
• relation — other named people in the user's life
• event — things that happened to the user (recent or ongoing)
• emotion — emotional state, current preoccupation

Each content is a third-person sentence, 3-30 words, with the user as implicit subject. Empty memories array is a valid output meaning "nothing notable in this window" — explicit training signal so the model learns when to stay silent rather than invent memories from chitchat.

Format

Each row is a chat-completion example in the OpenAI / HuggingFace messages convention:

{
  "seed_id": "...",
  "stratum": {
    "length": "medium", "density": "mixed", "register": "casual",
    "phase": "getting_to_know", "topic_mix": "scattered", "language": "en"
  },
  "messages": [
    {"role": "system", "content": "You are a memory-extraction model..."},
    {"role": "user", "content": "User: Nina\nFlame: Marisol\n\nConversation:..."},
    {"role": "assistant", "content": "{\"memories\":[...]}"}
  ]
}

Splits

• train.jsonl — 1728 examples
• eval.jsonl — 170 examples (~10%)

Stratified split bucketed on (density, language, phase) so the eval distribution mirrors train.

Density mix

Density	Train	Eval	Notes
mixed	738	77	Realistic average — facts mixed with banter
sparse	403	39	Mostly chitchat, ≤1 fact buried in flow
dense	363	34	Fact-heavy windows, 5+ user facts
null	224	20	Zero extractable user facts — empty array is the right answer

The 13% null share is deliberately above target (10%) — Phoenix needs strong "empty array" training to avoid inventing memories from pure banter.

Language mix

Heavily English (≥85%), with smaller stratified shares of Spanish, French, Japanese, and Chinese. Cross-lingual robustness, not full multilingual coverage. Phoenix outputs are always English regardless of input language.

Generation pipeline

Three stages, all reproducible from flammenai/Phoenix (in-house):

1. Seeds (Claude Haiku 4.5): synthetic conversations matching a stratum sampled from the axes above. ~2500 conversations generated.
2. Chosen extractions (Claude Sonnet 4.6): tool-forced JSON extraction with a 9-rule system prompt + 3 in-context examples covering null window, vulnerable-with-time-stripped, and ephemeral-filtered cases.
3. Quality judge (local Qwen3.5-27B + JSON-schema constraint): per- extraction verdict on an 11-mode failure taxonomy. Only extractions passing all 11 modes survive into this dataset.

Failure-mode taxonomy

The judge filters against these — every kept extraction passed all of them:

Mode	Description
hallucinated_date	Absolute date/year not in conversation
relative_time_qualifier	Phrases like "last month" that go stale
over_inference	Claim stronger than messages support
invented_from_null	Memories from a no-signal conversation
vague_filler	Generic "likes things"-tier content
flame_as_user	Captures Flame's words as user fact
ephemeral_detail	One-off mundane noise ("had pho today")
wrong_category	Category enum mismatch
wrong_language	Memory text not in English
merged_relations	Two distinct people collapsed
refusal_from_register	Skips memorable content from flirty/intimate windows

gold_anchors.json includes 10 hand-crafted reference examples with both chosen and rejected variants demonstrating these modes — useful for qualitative inspection.

Limitations

• SFT only. v1 is supervised fine-tuning. The DPO infrastructure (rejected-sample generator, per-mode contrast pairs) exists in the pipeline repo but stage 2b was not run at scale due to budget. v2 intends to add DPO refinement using real-world failure data collected from production extractions.
• Synthetic only. All conversations are Haiku-generated. Real-user conversation distributions will differ; production deployment should add real-data evals.
• Date awareness. Phoenix has no concept of "today" — relative time qualifiers are stripped on purpose. Memories like "Grandmother recently passed away" stay accurate even after months pass; "passed away last month" wouldn't.
• English-output bias. Memory content is always English even when the input conversation is in another language. By design — the consumer (FlameWorker prompt-builder) is English.
• Judge isn't perfect. The Qwen 27B judge reliably catches structural failure modes (vague_filler, wrong_language, hallucinated_ date, relative_time_qualifier) but is weaker on subtler semantic modes (invented_from_null, ephemeral_detail). Some chosens with subtle issues may have slipped through.

Contact

Schneewolf Labs LLC — flammen.ai