Commit History
4 days ago
-
Add a known runtime issue: the stop token with a zero embedding from the second turn
Verified signatureAdds the lines under Known runtime issues; nothing else in the card changes.a51f218
22 days ago
-
manifest 0.1.3: capabilities.thinking.control = never (what the bundle's template does about thinking, derived from the template inside the bundle)
Verified signaturef508c26
-
manifest: same-device CPU control rows on the Galaxy S26 for the flagship GPU recommendation (S7 backfill 2026-09-05); the recommendation is rewritten from the measured pair (measured win, kept for prefill, or reversed to cpu)
Verified signatured786967
-
Add litertlm_manifest.json (deployment manifest 0.1.2: variant selection, verified backends, Android/Apple recommendations, measured rows with conditions)
Verified signaturef44642c -
Card: base_model_relation: quantized (list under the base model's Quantizations)
Verified signatured62a802
-
Card: note the chat-template fix for role-'assistant' history (weights unchanged)
Verified signature0b5d240 -
Chat template fix (metadata-only, weights unchanged): the template matched history turns only on role 'model', so a conversation created through the Conversation API with the standard role 'assistant' (e.g. create_conversation(messages=[...]) restoring a chat) silently dropped every assistant turn from the rendered prompt. The role conditions now also accept 'assistant'. Every other section is byte-identical; single-turn renders, token ids, prefill and greedy answers are unchanged (verified on the LiteRT-LM runtime). See google-ai-edge/LiteRT-LM#3417.
Verified signature41b14fd
-
Card: note the tokenizer-section fix (weights unchanged)
Verified signaturebec2ce9 -
Replace the tokenizer section with the upstream tokenizer.json (weights, graph and metadata unchanged). The previous SentencePiece conversion of the BPE tokenizer encoded standalone accented/special characters to the wrong ids, turned characters without a vocabulary entry (emoji, some accented capitals) into the end-of-text/end-of-turn token, and could not match the token it had reused as UNK; prompts now tokenize identically to the upstream tokenizer (verified on the LiteRT-LM runtime, see the card note).
Verified signatureb7bd451
-
Card: correct the start_token note — measured numbers do not carry over
Verified signaturea60d0c0 -
Card: note the start_token fix (weights unchanged)
Verified signature5b8977d -
Drop the metadata start_token (weights unchanged)
Verified signatureThe metadata start_token held the literal string "None" (this tokenizer has no BOS), and the engine resolved it to a real vocabulary token - so every prompt began with the word "None", a stream the model was never trained on. Metadata-only change: every tokenizer and TFLite section is byte-identical to the previous file (verified by sha256 per section), so quality and speed are unchanged and the card's measured numbers still stand.aa75d39
-
Add measured Galaxy S26 GPU backend section
Verified signatureFunctional verdict from the S26 GPU gate. Primary log: litertlm-convert/community_accel_work/s4_gpu_gate/results.jsonl. Delegation counts and peak RSS are read from that run's own log; throughput is deliberately absent, because an Android GPU speed is not quotable without a CPU row from the same phone and that control has not been run.0583e29
-
Card: verified one-command usage, measured Performance table (backends gated on real generation), accuracy note
Verified signature52b8757
-
Upload model.litertlm with huggingface_hub
Verified signature9d649ed -
Upload README.md with huggingface_hub
Verified signature0f30487 -
initial commit
Verified signature111f6a9