elvin2929/functiongemma-270m-it-thinq-compat

FunctionGemma 270M IT — preformatted-prompt metadata compatibility experiment

Modified GGUF. Experimental, metadata-only derivative; not a fine-tuned model.

This repository contains one Q4_K_M GGUF derived from Google's FunctionGemma 270M IT. The immediate input was the Q4_K_M artifact distributed as version 1 of palm/functiongemma-270m-it on Melange. The original public model and its weights have not been modified in place. This is an independent compatibility experiment, not an official Google or LG model.

Intended use and limitations

The target application supplies a complete, already serialized FunctionGemma prompt as the last message's content, including the function declarations, developer and user turns, and open model turn. The passthrough template avoids serializing those turns a second time. It is not a normal conversational chat template: do not pass an ordinary multi-message conversation to this model and expect it to format the conversation or tools for you. Earlier messages are ignored by this template. BOS insertion remains tokenizer-owned.

The added EOT metadata marks <start_function_response> as an intended generation boundary after a function call. Whether a particular SDK/runtime honors this metadata remains to be tested. This does not add native function calling support to an SDK.

The target ThinQ demo executes tools through a mock adapter. This repository does not contain appliance profiles, app source, credentials, private prompts, device logs, or any real-appliance connection. It does not expand the model's training, supported appliance operations, or demonstrated accuracy.

Status at publication: structural integrity checks and host template rendering checks passed. Android/iOS SDK generation, stop behavior, function calling accuracy, Melange delivery, and Device Farm validation have not yet been established for this derivative. No runtime success or safety claim is made.

Exact metadata changes

Only these two stored GGUF metadata entries differ from the immediate input:

Key	New value	GGUF type
tokenizer.chat_template	{{ messages[-1]['content'] }}	STRING
tokenizer.ggml.eot_token_id	50 (<start_function_response>)	UINT32

The template has no trailing newline and emits no textual BOS. Existing add_bos_token=true, BOS ID 2, EOS ID 1, and add_eos_token=false are unchanged. All other stored metadata, token vocabulary, tensor types, shapes, and weights are preserved. The GGUF header's metadata-entry count necessarily increases by one when adding the previously absent EOT key.

Artifact integrity

File: functiongemma-270m-it-thinq-compat-Q4_K_M.gguf

• Format: GGUF v3; 236 tensors.
• File size: 253,113,408 bytes.
• File SHA-256: 35373153ddbb6ad6ede046d45efdfba49c1ab918a75f7adac83a6beb7250d19d.
• Immediate input size: 253,127,136 bytes.
• Immediate input SHA-256: 51c8c84d40dcf313482970bf8b3db34d11e61e9bdb3450314be5b3d647dc92ab.
• Ordered raw tensor payload SHA-256 (identical before/after): 79ff0e9b47749b89f38d7bd303c6c0248cadd56746d073641ca205308e785a59.

Verification compared every tensor payload byte and all non-target stored metadata. A separate raw GGUF parser independently checked the same invariants. This is evidence of preserved model data, not runtime equivalence or accuracy. Modification date: 2026-10-21.

License and required use restrictions

The model and this derivative are governed by the Gemma Terms of Use, reproduced in LICENSE, including Section 3.2 and the incorporated Gemma Prohibited Use Policy. Use and distribution of this derivative are subject to those restrictions; no additional grant overrides them. See also NOTICE. By using, modifying, or distributing the model, you are subject to the agreement as specified in that license. Google retains the rights in the upstream model.