# Llama shrinker v1 report

Transformers: `4.55.0`  
Allowed modified fields: `bos_token_id, eos_token_id, hidden_size, intermediate_size, max_position_embeddings, num_attention_heads, num_hidden_layers, num_key_value_heads, pad_token_id, vocab_size`

| Repository | Source pre-check | Config / model class | Original geometry | Shrunk geometry | Modified fields | Parameters | Checkpoint bytes | Post-check |
| --- | --- | --- | --- | --- | --- | ---: | ---: | --- |
| `TinyLlama/TinyLlama-1.1B-Chat-v1.0` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":2048,"intermediate_size":5632,"num_hidden_layers":22,"num_attention_heads":32,"num_key_value_heads":4,"vocab_size":32000,"max_position_embeddings":2048}` | `{"hidden_size":128,"intermediate_size":256,"num_hidden_layers":2,"num_attention_heads":16,"num_key_value_heads":2,"vocab_size":128,"max_position_embeddings":128}` | `hidden_size,intermediate_size,max_position_embeddings,num_attention_heads,num_hidden_layers,num_key_value_heads,vocab_size` | 303744 | 1217160 | POSTCHECK_OK |
| `HuggingFaceTB/SmolLM2-135M-Instruct` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":576,"intermediate_size":1536,"num_hidden_layers":30,"num_attention_heads":9,"num_key_value_heads":3,"vocab_size":49152,"max_position_embeddings":8192}` | `{"hidden_size":48,"intermediate_size":96,"num_hidden_layers":2,"num_attention_heads":6,"num_key_value_heads":2,"vocab_size":128,"max_position_embeddings":128}` | `hidden_size,intermediate_size,max_position_embeddings,num_attention_heads,num_hidden_layers,num_key_value_heads,vocab_size` | 46320 | 187328 | POSTCHECK_OK |
| `mtgv/MobileLLaMA-1.4B-Chat` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":2048,"intermediate_size":5632,"num_hidden_layers":24,"num_attention_heads":16,"num_key_value_heads":16,"vocab_size":32000,"max_position_embeddings":2048}` | `{"hidden_size":16,"intermediate_size":32,"num_hidden_layers":2,"num_attention_heads":2,"num_key_value_heads":2,"vocab_size":128,"max_position_embeddings":128}` | `hidden_size,intermediate_size,max_position_embeddings,num_attention_heads,num_hidden_layers,num_key_value_heads,vocab_size` | 9296 | 39288 | POSTCHECK_OK |
| `openbmb/MiniCPM5-1B` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":1536,"intermediate_size":4608,"num_hidden_layers":24,"num_attention_heads":16,"num_key_value_heads":2,"head_dim":128,"vocab_size":130560,"max_position_embeddings":131072}` | `{"hidden_size":128,"intermediate_size":256,"num_hidden_layers":2,"num_attention_heads":16,"num_key_value_heads":2,"head_dim":128,"vocab_size":128,"max_position_embeddings":128}` | `eos_token_id,hidden_size,intermediate_size,max_position_embeddings,num_hidden_layers,vocab_size` | 1409664 | 5640864 | POSTCHECK_OK |
| `deepseek-ai/deepseek-coder-6.7b-instruct` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":4096,"intermediate_size":11008,"num_hidden_layers":32,"num_attention_heads":32,"num_key_value_heads":32,"vocab_size":32256,"max_position_embeddings":16384}` | `{"hidden_size":16,"intermediate_size":32,"num_hidden_layers":2,"num_attention_heads":2,"num_key_value_heads":2,"vocab_size":128,"max_position_embeddings":128}` | `bos_token_id,eos_token_id,hidden_size,intermediate_size,max_position_embeddings,num_attention_heads,num_hidden_layers,num_key_value_heads,vocab_size` | 9296 | 39288 | POSTCHECK_OK |
| `NousResearch/Hermes-3-Llama-3.1-8B` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":4096,"intermediate_size":14336,"num_hidden_layers":32,"num_attention_heads":32,"num_key_value_heads":8,"vocab_size":128256,"max_position_embeddings":131072}` | `{"hidden_size":64,"intermediate_size":128,"num_hidden_layers":2,"num_attention_heads":8,"num_key_value_heads":2,"vocab_size":128,"max_position_embeddings":128}` | `bos_token_id,eos_token_id,hidden_size,intermediate_size,max_position_embeddings,num_attention_heads,num_hidden_layers,num_key_value_heads,vocab_size` | 86336 | 347488 | POSTCHECK_OK |
| `livekit/turn-detector` | SOURCE_OK | LlamaConfig / LlamaForCausalLM | `{"hidden_size":576,"intermediate_size":1536,"num_hidden_layers":30,"num_attention_heads":9,"num_key_value_heads":3,"head_dim":64,"vocab_size":49154,"max_position_embeddings":8192}` | `{"hidden_size":48,"intermediate_size":96,"num_hidden_layers":2,"num_attention_heads":6,"num_key_value_heads":2,"head_dim":64,"vocab_size":128,"max_position_embeddings":128}` | `hidden_size,intermediate_size,max_position_embeddings,num_attention_heads,num_hidden_layers,num_key_value_heads,vocab_size` | 132336 | 531400 | POSTCHECK_OK |

## Per-source provenance

- `TinyLlama/TinyLlama-1.1B-Chat-v1.0`: revision `fe8a4ea1ffedaf415f4da2f062534de366a451e6`, source SHA-256 `486bedda3a6988332e60d9638a09ca4b260d34ebcf1b19e22cf3b140b63d8fe9`, rules `gqa-ratio-8-preserved`.

- `HuggingFaceTB/SmolLM2-135M-Instruct`: revision `12fd25f77366fa6b3b4b768ec3050bf629380bac`, source SHA-256 `8eb740e8bbe4cff95ea7b4588d17a2432deb16e8075bc5828ff7ba9be94d982a`, rules `gqa-ratio-3-preserved`.

- `mtgv/MobileLLaMA-1.4B-Chat`: revision `7073081bd6b02d97ac4f9c7e8167fd3c28483c1c`, source SHA-256 `1e0f3e66ba7c3824ae2a2352df692c50d0e1b80df3b1fd1338670084de905aaf`, rules `mha-preserved`.

- `openbmb/MiniCPM5-1B`: revision `4e9de7a0778dc1c362e983e6858f0e77542cbdca`, source SHA-256 `6a6509b646cb3169616c5ffc3196e7ccaf9d4d6bc17b266581d241a31c217714`, rules `gqa-ratio-8-preserved, out-of-range-special-tokens-remapped, explicit-head-dim-preserved`.

- `deepseek-ai/deepseek-coder-6.7b-instruct`: revision `e5d64addd26a6a1db0f9b863abf6ee3141936807`, source SHA-256 `a3a2ce1a623b367d3799fb0c1e1ba12d5b8fa51a8e7abba14d10dd3739cd73e4`, rules `mha-preserved, out-of-range-special-tokens-remapped, rope-scaling-preserved`.

- `NousResearch/Hermes-3-Llama-3.1-8B`: revision `896ea440e5a9e6070e3d8a2774daf2b481ab425b`, source SHA-256 `7e189a0da074da6a878d5a276336caf3d62c57138f461bc6239402639f1bc6f5`, rules `gqa-ratio-4-preserved, out-of-range-special-tokens-remapped, rope-scaling-preserved`.

- `livekit/turn-detector`: revision `fba34c38ad5d30a63ebb83a9e6bf271cf4c91d67`, source SHA-256 `e8feb37fa6f974232742d4e223e40928ccc8dabe1906ce97090e14b5753dbcc1`, rules `gqa-ratio-3-preserved, explicit-head-dim-preserved`.

## Global findings

One repository-independent shrinker handled 7/7 source configs that passed pre-check.

Dependency rules: query heads remain divisible by KV heads; MHA equality is preserved; GQA keeps its observed query/KV ratio and at least two KV heads; hidden width is eight values per reduced query head; intermediate width is twice hidden width; context is reduced to 128 positions; only out-of-range special IDs are remapped while preserving representation and equality; explicit `head_dim`, both observed `rope_scaling` forms, and semantic bias/tie fields are preserved.

Special treatment was required only for explicit `head_dim` preservation and deterministic special-token remapping. Both the linear and Llama-3 RoPE objects execute without dependent changes after context is reduced to 128. Transformers emits a non-fatal consistency warning for the preserved Llama-3 `original_max_position_embeddings=8192`, which is recorded as a validation observation.

Remaining v0 limitation: MQA and unobserved special-token field names are intentionally not generalized from this seven-config corpus.