CLM-v0.1-8B projection heads in GGUF
This repository contains the state and action projection heads only from
Contrastive-LM/CLM-v0.1-8B.
CLM scores candidate actions against a state; it does not generate text. A separate,
frozen Qwen3-8B encoder is required to turn text into embeddings.
| File |
Contents |
Size |
SHA-256 |
clm-v0.1-8B-heads-f32.gguf |
17 F32 tensors, including both projection heads |
75,552,224 bytes |
d5bbef8f4c9edf0b5cfb38b4a9326a1ac776cc9b479345a56b2efc37de3c39be |
The source checkpoint is CLM_v0.1-8B.pt (SHA-256
b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5).
The converter
preserves the F32 weight and bias bytes. It stores the reference scoring scale
min(exp(logit_scale), 100) = 100 in the GGUF. The GGUF does not contain
the Qwen3-8B encoder, tokenizer, or original .pt file.
Download and run
The verified encoder file is
unsloth/Qwen3-8B-GGUF / Qwen3-8B-BF16.gguf,
SHA-256 5e416a2020fe63e76ea13c8979be35fc6070aaf3578f7876400c55c2f5c3eb30.
Use rust-model-inference
at or after the merged CLM parity change #128,
then download both files:
git clone https://github.com/Liyulingyue/rust-model-inference.git
cd rust-model-inference
hf download EvoAwaken-Workshop/CLM-v0.1-8B-gguf \
clm-v0.1-8B-heads-f32.gguf --local-dir models/CLM-v0.1-8B
hf download unsloth/Qwen3-8B-GGUF Qwen3-8B-BF16.gguf \
--revision a6adef130ffb23ddaf1a62fec9dced968c9bc482 \
--local-dir models/Qwen3-8B-GGUF
CLM uses the existing --jev candidate-scoring interface:
cargo build --profile release-fast --bin rust-model-inference
./target/release-fast/rust-model-inference \
--model models/Qwen3-8B-GGUF/Qwen3-8B-BF16.gguf \
--jev --clm-head models/CLM-v0.1-8B/clm-v0.1-8B-heads-f32.gguf \
--jev-context 'Customer: my invoice was charged twice.' \
--jev-question 'Which team should handle this?' \
--jev-option 'Charges, invoices, refunds' \
--jev-option 'Bugs and outages' \
--threads 1
The state input is context + "\n\n" + question; each candidate is encoded as
written. The encoder uses last-token pooling and L2 normalization before the
two heads. Scores and probabilities are relative to the supplied candidates.
Conversion and verification
The official CLM implementation defines
the two heads and scoring rule. The GGUF uses general.architecture=clm and is
loaded alongside a qwen3 encoder. The source checkpoint and converter pass a
tensor round-trip check.
On 2026-09-30, two real-text scoring cases matched a fixed scalar llama.cpp
encoder and an independent scalar CLM-head implementation: token IDs, 32,768
encoder F32 values, and 34,818 projection-head F32 values were bitwise
identical. Final score bit patterns were 0x415a1e78 and 0x41b29eb4.
Hashes, commands, and the verification script
record the exact scope.
This bitwise result applies to the specified BF16 encoder and F32 heads in
scalar mode. The official vLLM service, NEON/FMA/BLAS/Accelerate, the default
accelerated path, and other quantizations have not been compared bitwise.
Attribution and license
The CLM heads and Qwen3-8B base model are released under Apache 2.0. See the
LICENSE file in this repository and the original model pages for details.
